ESSAYS ON THE VOLUME, ISSUE SCOPE, AND INTER-NATION ALIGNMENT OF

UNITED NATIONS GENERAL ASSEMBLY SPEECHES

By

Zuhaib Sheikh Mahmood

A DISSERTATION

Michigan State University

in partial fulﬁllment of the requirements

Submitted to

for the degree of

Political Science – Doctor of Philosophy

2020

PUBLIC ABSTRACT

ESSAYS ON THE VOLUME, ISSUE SCOPE, AND INTER-NATION ALIGNMENT OF

UNITED NATIONS GENERAL ASSEMBLY SPEECHES

By

Zuhaib Sheikh Mahmood

A range of literature has argued that countries use the United Nations (UN) to communicate their
interests and coordinate their foreign policies. However, a core behavior at the UN – public-facing
speeches – have remained understudied. While negotiation often takes place in private, the puzzle
remains of why countries would then turn to a public forum to communicate–given the always-
present option of a private one. In this dissertation, I expect that public forums serve a distinct
purpose as compared to private negotiations: rather than being persuasive in nature, these public
communications serve as being political in nature. I expect that the countries who beneﬁt most
from this broadcasting capability will be those who beneﬁt most from communicating their position
to a large audience, be those audiences international or domestic. In order to isolate the purely
political nature of these public-facing speeches, I draw on the text of speeches delivered at the
United Nations General Assembly (UNGA) for the years 1984-2014.

ABSTRACT

ESSAYS ON THE VOLUME, ISSUE SCOPE, AND INTER-NATION ALIGNMENT OF

UNITED NATIONS GENERAL ASSEMBLY SPEECHES

By

Zuhaib Sheikh Mahmood

A range of literature has argued that countries use the United Nations (UN) to communicate their
interests and coordinate their foreign policies. However, a core behavior at the UN – public-facing
speeches – have remained understudied. While negotiation often takes place in private, the puzzle
remains of why countries would then turn to a public forum to communicate–given the always-
present option of a private one. In this dissertation, I expect that public forums serve a distinct
purpose as compared to private negotiations: rather than being persuasive in nature, these public
communications serve as being political in nature. I expect that the countries who beneﬁt most
from this broadcasting capability will be those who beneﬁt most from communicating their position
to a large audience, be those audiences international or domestic. In order to isolate the purely
political nature of these public-facing speeches, I draw on the text of speeches delivered at the
United Nations General Assembly (UNGA) for the years 1984-2014.

This dissertation is composed of three chapters, and makes two broad contributions. First, it
provides political grounding to one of the most public, visible, and proliﬁc behaviors in interna-
tional politics: public-facing political speeches delivered at international institutions. This has
implications for both the study of the UN and the study of diplomacy more broadly, and provides
a framework to study other institutions in the future. Second, it casts doubt on the conventional
wisdom that these types of speeches are not informative:
to the contrary, I show that all three
dimensions of speech–volume, content, and alignment–map on to patterns of behavior–namely,
UN Peacekeeping troop deployment; intervention into civil wars; and UNGA voting similarity.
This suggests that each dimension of political speech contains useful information about countries’
underlying interests.

In the ﬁrst chapter, I analyze why countries engage with the UN General Assembly at all. I
argue that engagement with the UNGA is driven by the political beneﬁts a country perceives from
communicating opinions to politically relevant audiences. I show that countries with the capacity
and interest to act in many issues, as well as countries with strong political incentives linking them
to inﬂuential audiences at home, will tend to speak the most at the UNGA. I also validate that the
volume of UNGA speech does in fact measure a perception of political beneﬁt from the UN more
broadly, showing that the amount of speech at the UNGA is strongly related to troop contributions
to UN Peacekeeping missions

In the second chapter, I examine the scope of interests that countries speak on in their public
facing speech. This provides, in part, a mechanism for the ﬁrst chapter: countries with the material
capacity to act in more issues, and countries with the political incentives to communicate with a
more diverse audience, will tend to speak on a more diverse set of issues in their speeches. I also
validate that the issues countries speak on tend to be ones of of political interest: I do so by showing
that when a country experiences a civil war, countries who speak similarly will be more likely to
intervene in some way in the conﬂict.

Finally, in the third chapter, I examine the conditions under which countries use speech to
reinforce or hedge against the underlying incentives and interests driving their behavior. I examine
two such conditions. First, I show under conditions where coercion is more likely–such as where
a large power diﬀerence exists–hedging is more likely since it is more likely there is a diﬀerence
between underlying positions and observed positions. Second, I show that conditions under which
friends and enemies are more clearly deﬁned–including periods of rivalry, joint democracy, and as
issues become more controversial–the cost of hedging will increase and thus countries will tend
to reinforce in their speech. I use UNGA voting alignment as behavioral baselines for measuring
reinforcement and hedging across issues. This chapter also relies on an ideal point measure–
created in collaboration with Michael Colaresi, using a Dynamic Linear Model to cast ideal points
in multiple dimensions based on speeches–to measure hedging and reinforcing speech.

Copyright by
ZUHAIB SHEIKH MAHMOOD
2020

To my father, Sheikh Tahir Mahmood,

for motivating me to always push forward

and

To my mother, Nuzhat Mahmood,
for giving me the strength to do so.

v

ACKNOWLEDGEMENTS

For the past seven and a half years, I have read multiple dissertations, each with its own set of
acknowledgements. Each time, I found myself thinking about the day I would ﬁnally get to write my
own. To properly acknowledge all the people who brought me to this point, and to put into writing
something to memorialize their contributions. There is a certain irony to the fact that today, when
the time has ﬁnally come to write this, I ﬁnd myself at a loss for words. Nearly every dissertation
before me has mentioned the diﬃculty of acknowledging those who deserve acknowledgement.
While I always thought I understood this intellectually, most things in life are as my mother and
father always told me: you will learn when you ﬁnally get there. I ﬁnally understand.

It’s only ﬁtting, then, that the greatest acknowledgement goes to my mother, Nuzhat Mahmood,
and my father, Sheikh Tahir Mahmood. In many ways, my life up to this point is analogous to
this dissertation: a stumbling, incomplete, and pretty underwhelming attempt to build upon the
great ones before me to reach new heights. Perhaps I’ve passed the bar for doing alright so far
in life, in the same way that this dissertation might pass the bar to continue in this ﬁeld, but to
have gotten this far is only to have gained the humility in realizing how far I have left to go. My
mother and father’s lives, building on their mothers, fathers, and elders before them, have achieved
new heights: leaving a lasting impact on a community, leaving a legacy to build upon for the next
generation, and teaching innovating new and important ways to live life. Through their innovations
and their work, I’ve been given the luxury in life to reach heights I never thought possible. While
every time I reach a bit higher I eventually lose my balance–inevitably falling over again–it’s only
through their achievements that I’m able to reach just high enough to maybe ﬁnd something new. I
can only hope that one day, in reaching higher and higher, I might ﬁnd something to grasp, and give
my shoulders to someone after me to reach even higher. Anything great in my life is only possible
because I stood on their shoulders to ﬁnd it.

Next, I have to thank Carina Saavedra. While I lived the fool’s life of continuously reaching
high, falling over, only to reach and fall again, she was able to provide me the inspiration of the

vi

grounded path. Where I might have just continued to fall back to zero, her support help raise the
ground beneath me, allowing me to fall slightly higher each time. When I needed to take a full time
job in the middle of graduate school, she was there to march forward while I worked until midnight
each day. When I was mentally exhausted after falling down yet again, she was there with her feet
planted, ready to help me up. And when life got diﬃcult because I felt like I was juggling too many
things, she was there to remind me that there was a ﬂoor beneath me, and that sometimes it’s OK
to put things down and pick them back up again. Her groundedness in life is only matched by her
genuine love for the Earth; much like her uncanny ability to nurture plant life to new and amazing
heights by nurturing the ground beneath her, she has a similar capacity to ground those around
her–however stubbornly they insist on having their heads in the clouds. For someone like me, who
sometimes (and with concerning ease) forgets which way is up or down, it’s hard to overstate how
valuable this really has been.

It goes without saying that it’s impossible to fully thank everyone else who deserves to be
thanked. But I suppose I can try. Katie M. Willett, a history teacher at Independence High School
and the Academy of International Studies program, was the ﬁrst person to inspire me to pursue and
research international politics. University of North Carolina (UNC) Political Science Professor Dr.
Mark Crescenzi was my main source of inspiration, and was always there for me as I ﬁne-tuned
my introduction to Political Science as a discipline. UNC Chemistry professor Dr. Domenic
Tiani–who quite literally could not contain his laughter when he saw my grades, when I said I
was considering medical school (even as he wrote me a letter of recommendation regardless)–
left a lasting impression on me as embodying the value of having incredibly honest, smart, and
sincere mentors around you. And, ﬁnally, I look back at all my teachers from Elementary through
University. I’ve had the luxury in life of remembering–fondly–every single teacher I have had since
Kindergarten. This may have been the single greatest privilege of my life: to have always had the
honor of being taught by passionate, memorable, and truly inspirational teachers at every single
step of my life. This is not a luxury are all granted in life, and for this I am truly thankful.

I should also thank my brothers, Atif Mahmood and Raheel Mahmood, for being the best

vii

brothers anyone could ask for. I’d say more, but they’re my brothers, which speaks for itself. They
spent a lot of time asking me when I would ﬁnish, and amused themselves with the fact that I’ve
spent 30 years in school to not even become a real doctor, and to do something so obscure where
most people wouldn’t care about it, or understand it. Naturally, this only motivated me more: after
all, what are brothers good for, if not to make you more stubborn? And when the time ﬁnally came,
as always, they were my biggest supporters and cheerleaders. I sincerely love them both.

My dissertation chair, Dr. Benjamin Appel, took me on as a project from the beginning. He
reached out to me after I was ﬁrst accepted to Michigan State as a graduate student, expressing
excitement at our shared interest in the U.N. He helped me to write my ﬁrst publishable paper
in graduate school, and he took me on as a dissertation chair after life threw curveballs midway
through graduate school–the way life tends to do. He read multiple drafts of these essays, and
showed both incredible patience and a truly constructive spirit by moving me forward on the path
to actually ﬁnishing this project. For anyone who reads this dissertation, you will quickly see why
this was such a blessing: the amount of work left to be done here is still incredibly high, but his
focused mentorship drove me to a place where I could at least stand ﬁrmly in what I had–rather than
drowning in what I could have done. To him, I can only say thank you for believing in me, for taking
me on, and for helping me through the ﬁnal stretch with an eye on helping me transition to my
future state–rather than trying to perfect me in my present state. They say that the best dissertation
is a completed dissertation, and without your guidance in this process, there’s a real chance I would
have continued chasing a better dissertation, perhaps sacriﬁcing the best dissertation for it. As I
edit, revise, and most likely entirely revamp this dissertation in the future, it will have been your
decision to help me stick the landing that allows this project to evolve to wherever it can go.

As I get closer to the end of these acknowledgements, I can only start listing out those who come
to mind, with apologies to everyone I inevitably forget. Dr. Matthew Grossman took me on as a
Research Assistant during a transitionary period in graduate school, leading to a much-needed reset,
a blog post publication, and two co-authored papers (including one publication). Working with him
has been one of the most unexpectedly productive, interesting, and educational opportunity I’ve

viii

had in graduate school, and I hope it continues moving forward. Dr. Robert Trager, who jumped
on my committee after I cornered him at a conference, has been the source of some deeply thought-
provoking inspiration on the topics of speech and diplomacy; I know I’ve only scratched the surface
the depth of knowledge I can gain from continuing to work with him in the future. More broadly,
I owe Michigan State University, and the Department of Political Science in particular, a massive
debt. The fact that this debt is not ﬁnancial only makes that debt deeper; with their support for
conferences, living, and for professional enrichment, I was able to complete this journey prepared
for the next step. A world of thanks to all the department staﬀ, including (but not limited to) Karen
Battin, Sarah Krause and Rhonda Burns. And, of course, I can’t go without acknowledging the
colleagues and close friends I’ve developed along the way, including Daniel Fram, Douglas Walker,
Jessica Schoenherr, and Peter Penar. Thank you for allowing me into your lives, and allowing me
to share my life with you for the short time we spent in East Lansing together.

At long last, I end with my biggest debt from graduate school. When I think about building
this house of cards (note: yes, I know what this means. It does feel this way most of the time)
I’ve built for myself, reaching high, falling down, and reaching again–only to fall on my face yet
again–I can think of no one who has watched me do this quite as much as Dr. Michael Colaresi.
When I ﬁrst applied to Michigan State University, everyone I spoke to said the same thing: if you
go to Michigan State, be sure to work with Michael Colaresi. He lives up to the legend. There
is no one in my life who I simultaneously respect and fear the way I do Dr. Colaresi, and there
is no one with whom I feel as inspired as I do every time I talk to him. In addition to his depth
of intelligence, his love for his work has been–by itself–one of his greatest gifts to me. I once ran
into him at a conference, where he said with some happiness that he’d brought his wife and kids
with him to make a small vacation out of the trip. In a ﬁeld where work-life balance and mental
health is a real issue, this was one of the most impactful moments of my time in graduate school.
Dr. Colaresi is someone who truly embodies every great trait I could hope for in an advisor and a
mentor: inspiring, humbling, and frustratingly impossible to live up to. He’s someone who I ﬁnd
myself always wanting to impress, always failing to do so, and just wanting to try harder next time.

ix

The depth of Dr. Colaresi’s intelligence is matched only by the depth of his patience, as he
oversaw my futile attempts to grasp whatever I could from our conversations, courses, and time
together. Our long conversations are the reason I know anything about this ﬁeld, and his tireless
commitment to push his students to do–no, to be–better is the reason I can actually do anything
useful in this ﬁeld. When he saw my laptop screen in a coﬀee shop one evening, color-coding a
Microsoft Excel document and trying to organize my data, he chastised me almost with incredulity,
insisting that this was not a sustainable way to manage data if I was going to do the type of work I
wanted to do. When I wanted to collect a large corpus of UN speeches, he gave me room to bang
my head against the wall for a year as I learned Python to get it done. And, at every step along
the way, he gave me opportunities: the opportunity to co-author two papers with him, including
one which became my ﬁrst publication, and the opportunity to spend time developing a project I
didn’t quite understand yet. Under his umbrella, for four incredible years, I was able to ﬂourish.
While life circumstances made it diﬃcult to stick the landing at the end, his mentorship is the only
reason I feel like I still have a ceiling above me in this ﬁeld, that I have yet to reach. To him, I can
do nothing else but apologize: while I’ve stumbled along the way, and I’m not where you probably
hoped I would be, I hope one day to reach the ceiling you spent so long building above me.

x

TABLE OF CONTENTS

.
.
.

.
.
.

.
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
. xviii
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
LIST OF TABLES .
.
LIST OF FIGURES .
KEY TO ABBREVIATIONS .

.
.

.
.

.
.

.
.

CHAPTER 1

1.1

1.2 Data collection .

.

.

.

.

.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Introduction .
1.1.1 The United Nations as a domain of analysis for speeches

INTRODUCTION TO THE DISSERTATION AND SUMMARY OF
DATA COLLECTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1
.
3
. . . . . . . . . .
8
.
8
Step 1: Scraping and converting to text . . . . . . . . . . . . . . . . . . . .
1.2.1.1 Conversion to text
. . . . . . . . . . . . . . . . . . . . . . . . . 11
Step 2: Parsing and Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.2.1 Cleaning: parsing error correction . . . . . . . . . . . . . . . . . 17
. . . . . . . . . . . 19
Step 3: Identifying speeches with individual countries

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2.2

1.2.3

1.2.1

.

.

.

.

.

CHAPTER 2 MATERIAL CAPACITY AND POLITICAL INCENTIVES AS DRIVERS

.

.

Variables of interest

2.2 Data and measurement

2.2.1 Measurement of country-level characteristics

OF UN GENERAL ASSEMBLY SPEECHES . . . . . . . . . . . . . . . . 28
2.1 The UN General Assembly as a platform for connecting speakers to audiences . . . 30
2.1.1 Material capacity, Political institutions, and UNGA speech . . . . . . . . . 32
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
. . . . . . . . . . . . . . . . 36
2.2.1.1
. . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.1.2 Control variables . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.1.3
. . . . . . . . . . . . . . . . . . . . . . 43
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3.1 Evaluating mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.3.1.1 Material capacity . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Political incentives . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.3.1.2
2.4 Linking speech to interest on relevant UN issues . . . . . . . . . . . . . . . . . . . 58
2.5 Discussion and conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2.3 Empirical patterns .

The dependent variable

.

.

.

.

CHAPTER 3 COMPOSITIONAL SPECIFICITY AND DIFFUSENESS OF SPEECHES

3.1.1

ACROSS WORLD ISSUES . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.1 Diplomacy, communication, and the audiences who listen . . . . . . . . . . . . . . 68
Statements of opinion and the UN General Assembly . . . . . . . . . . . . 70
Empirical implications . . . . . . . . . . . . . . . . . . . . . . . 71
3.1.1.1
3.2 Data and issue composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
. 76
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

3.2.1
3.2.2 Model and variables
.

Issue diﬀuseness and speciﬁcity . . . . . . . . . . . . . . . . . . . . . .

3.3 Analysis .

.

.

.

.

.

.

.

.

xi

3.3.1 Alternative measures and the content of speeches . . . . . . . . . . . . . . 86
Resolution topics . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3.1.1
3.3.1.2 Dirichlet Regression . . . . . . . . . . . . . . . . . . . . . . . . 87
. 96
. 99

3.4 Does issue prioritization correspond to issue salience? . . . . . . . . . . . . . . .
3.5 Discussion and conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 4 COMMUNICATING FOREIGN POLICY ALIGNMENT VIA STATE-
MENTS OF OPINION AT THE UNITED NATIONS GENERAL AS-
SEMBLY .

.

.

.

.

4.2.1.1

4.3 Empirical strategy .

4.2.1 Mapping speech to friends and foes

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.1 Speech, foreign policy alignment, and the United Nations . . . . . . . . . . . . .
. 103
4.2 A scale of two interests: Alignment through relative content . . . . . . . . . . . . . 107
. . . . . . . . . . . . . . . . . . . . . 109
Empirical implications . . . . . . . . . . . . . . . . . . . . . . . 110
. 114
4.3.1 Mapping speeches to resolutions . . . . . . . . . . . . . . . . . . . . . . . 114
4.3.1.1 Measuring alignment of speech . . . . . . . . . . . . . . . . . . 117
4.3.1.2 Measuring levels of controversy . . . . . . . . . . . . . . . . . . 118
Behavioral similarity . . . . . . . . . . . . . . . . . . . . . . . . 119
4.3.1.3
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4.3.2 Research design .
.

4.4 Results and evaluation .

.

BIBLIOGRAPHY .

.

.

.

.

.

.

.

.

.

.

.

xii

LIST OF TABLES

Table 1.1: Groups found in UN General Assembly speeches . . . . . . . . . . . . . . . . . 22

Table 2.1: Bivariate correlations with UNGA Word Count

. . . . . . . . . . . . . . . . . . 47

Table 2.2: Negative Binomial Regression, UNGA word counts, yearly . . . . . . . . . . . . 50

Table 2.3:

Incident Response Ratios for variables of interest

. . . . . . . . . . . . . . . . . 51

Table 2.4: UNGA Word Count (mean-centered), with Power*IGO interactions

. . . . . . . 56

Table 2.5: UNGA Word Count, with V-Dem*IGO interactions . . . . . . . . . . . . . . . . 59

Table 2.6: Replication of Kathman and Melin (2017) . . . . . . . . . . . . . . . . . . . . . 63

Table 3.1: Variable correlations .

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Table 3.2: OLS Regressions .

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

Table 3.3: OLS Regression - Military Expenditure . . . . . . . . . . . . . . . . . . . . . . 85

Table 3.4: Beta regression results, proportion of Plenary meetings; Resolutions; or

Agenda items connected to a country speaking (precision model results omitted)

88

Table 3.5: Collapsed Topics (1-5) .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

Table 3.6: Collapsed Topics (6-10)

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Table 3.7: Collapsed Topics (11-15) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Table 3.8: Collapsed Topics (16-20) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

Table 3.9: GLM models: Conﬂict involvement

. . . . . . . . . . . . . . . . . . . . . . . . 100

Table 4.1: 31-category composition for UNGA resolution A/RES/47/214 . . . . . . . . . . 116

Table 4.2: Binomial model of dyadic disagreement per topic; interaction with power diﬀerence125

Table 4.3: Binomial model of dyadic disagreement per topic; interaction with Rivalry . . . 127

Table 4.4: Binomial model of dyadic disagreement per topic; interaction with Joint Democracy128

xiii

Table 4.5: Binomial model of dyadic disagreement per topic; interaction with issue Con-

. .

. .

troversy .

.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

xiv

LIST OF FIGURES

Figure 1.1: HTML for PDF ﬁle on UN website . . . . . . . . . . . . . . . . . . . . . . .

. 11

Figure 1.2: Pre-Session-48 formatting of PDF transcript

. . . . . . . . . . . . . . . . . .

. 12

Figure 1.3: Post-Session-47 formatting of PDF transcript . . . . . . . . . . . . . . . . . . . 12

Figure 1.4: Mac Folder Action for ABBYY FineReader

. . . . . . . . . . . . . . . . . . . 13

Figure 1.5: Machine readable text ﬁle with speeches highlighted . . . . . . . . . . . . . . . 13

Figure 1.6: Example speakers with parenthetical identiﬁcation . . . . . . . . . . . . . . . . 19

Figure 1.7: Example speaker without identiﬁcation . . . . . . . . . . . . . . . . . . . . .

. 20

Figure 1.8: CSV ﬁle rows with speakers and speeches

. . . . . . . . . . . . . . . . . . . . 20

Figure 1.9: Total word count over time

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. 26

Figure 1.10: Mean word count per speech, over time . . . . . . . . . . . . . . . . . . . . . . 27

Figure 2.1: Total word count .

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Figure 2.2: Average yearly word count

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. 37

Figure 2.3: Average yearly word count (NAM)

. . . . . . . . . . . . . . . . . . . . . . .

. 38

Figure 2.4: Average yearly word count (EU)

. . . . . . . . . . . . . . . . . . . . . . . .

. 39

Figure 2.5: Average yearly word count (AU)

. . . . . . . . . . . . . . . . . . . . . . . . . 39

Figure 2.6: Average yearly word count per country . . . . . . . . . . . . . . . . . . . . . . 44

Figure 2.7: UNGA word count vs CINC . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 48

Figure 2.8: UNGA word count vs Democracy . . . . . . . . . . . . . . . . . . . . . . . . . 49

Figure 2.9: Predicted word count (with 95% CI), UNGA word count (cent) vs Nuclear

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

capability .

.

. .

Figure 2.10: Predicted word count (with 95% CI), UNGA word count (cent) vs CINC . . .

. 51

xv

Figure 2.11: Predicted word count (with 95% CI), UNGA word count (cent) vs VDem

Polyarchy .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

Figure 2.12: Predicted word count (with 95% CI), UNGA word count (cent) vs VDem

Participatory .

. .

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

Figure 2.13: Marginal eﬀect of CINC score across IGO membership, and vice versa . . . . . 55

Figure 2.14: Marginal eﬀect of Log GDP Per Capita across IGO membership, and vice versa

57

Figure 2.15: 2D perspective plot, UNGA word count vs Democracy and IGO membership . . 58

Figure 2.16: Marginal eﬀects (with 95% CI), vDEM Particip and IGO membership . . . . . 60

Figure 2.17: Loess smooth bivariate relationship, UNGA engagement and troop contribution

62

Figure 2.18: Predicted Counts of UNPK troops across word count

. . . . . . . . . . . . . . 62

Figure 3.1: Snapshot of UN General Assembly transcript . . . . . . . . . . . . . . . . . . . 75

Figure 3.2: Distribution of maximum ALR value of year-aggregated speeches, 1984-2014 . 79

Figure 3.3: Descriptive relationship between measures of power and issue speciﬁcity . . . . 81

Figure 3.4: Descriptive relationship between audience size and issue speciﬁcity . . . . . . . 82

Figure 3.5: Predicted max ALR value across Democracy Score . . . . . . . . . . . . . . . 84

Figure 3.6: Predicted max ALR value across log-Population . . . . . . . . . . . . . . . . . 84

Figure 3.7: Snapshot of UN General Assembly Resolution metadata . . . . . . . . . . . . . 86

Figure 3.8: Predicted composition across categories, across GDP Per Capita

. . . . . . . . 95

Figure 3.9: Predicted composition across categories, across Democracy . . . . . . . . . . . 95

Figure 3.10: Descriptive relationship between Euclidean distance in speech composition

and civil war involvement

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

Figure 3.11: Predicted probability of involvement, across values of Euclidean Distance

. . . 99

Figure 4.1: Three countries speaking in similar ideological spaces . . . . . . . . . . . . . . 107

Figure 4.2: Three countries speaking in very diﬀerent ideological spaces

. . . . . . . . . . 107

xvi

Figure 4.3: Number of resolutions, by topic . . . . . . . . . . . . . . . . . . . . . . . . . . 119

Figure 4.4: Speech alignment, by topic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

Figure 4.5: Controversy level, by topic

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. 120

Figure 4.6: Change in linear predictor of Speech diﬀerence vs. Disagreement, across

levels of power diﬀerence (95% conﬁdence intervals)

. . . . . . . . . . . . . . 126

Figure 4.7: Change in linear predictor of Speech diﬀerence vs. Disagreement, across

levels of rivalry (95% conﬁdence intervals) . . . . . . . . . . . . . . . . . . . . 129

Figure 4.8: Change in linear predictor of Speech diﬀerence vs. Democracy, across levels

of joint Democracy (95% conﬁdence intervals) . . . . . . . . . . . . . . . . . . 130

Figure 4.9: Change in linear predictor of Speech diﬀerence vs. Disagreement, across

levels of controversy (95% conﬁdence intervals) . . . . . . . . . . . . . . . . . 132

xvii

KEY TO ABBREVIATIONS

UN United Nations
UNGA United Nations General Assembly
PDF Portable Document Format
ODS Oﬃcial Document System
CSV Comma-Separated Values
COW Correlates of War

xviii

INTRODUCTION TO THE DISSERTATION AND SUMMARY OF DATA COLLECTION

CHAPTER 1

1.1 Introduction

In June of 2013, Hassan Rouhani was elected president of the Islamic Republic of Iran, winning
an overwhelming 50.71% of the vote–ahead of the runner-up Mohammad Baqer Qalibaf, who won
about 16.56% of the vote.1 This prompted immediate reactions from both within Iran and outside
of it. With “hard-liners” within his own country, Rouhani faced criticism over perceptions that
he would capitulate to Western pressure on the nuclear issue,2 and many Western elites remained
cautiously optimistic about the prospects for diplomatic opportunities with Iran that were not
possible with Rouhani’s predecessor, Mahmoud Ahmadinejad3-including a “historic” phone call
with U.S. president Barak Obama, the ﬁrst direct talks between the United States and Iran since the
1979 revolution.4

This strong shift in tone–what The Guardian termed a “diplomatic oﬀensive”–by Rouhani also
caused concern for Israel’s hawkish5 Prime Minister, Benjamin Netanyahu, raising concerns both
of a future detente between the U.S. and Iran, and concerns that the shift simply reﬂected “a wolf
in sheep’s clothing” on the part of Rouhani.6 Ultimately, while only time will reveal whether the
Rouhani regime signaled a structural shift toward the interests of Iranian reformists and away from
more hard-line, hawkish conservative interests,7 the reactions to Rouhani’s moderate rhetoric–

1http://www.bbc.com/news/world-middle-east-22916174
2https://foreignaﬀairs.org/articles/iran/2013-06-16/why-rouhani-won-and-why-khamenei-let-

3http://www.bbc.com/news/world-middle-east-22916174
4https://www.theguardian.com/world/2013/sep/27/obama-phone-call-iranian-president-

5https://www.theguardian.com/world/2013/sep/29/netanyahu-rouhani-iran-rhetoric
6http://www.washingtontimes.com/news/2013/oct/1/israels-benjamin-netanyahu-new-iran-

him

rouhani

leader-hassan-/

him

7https://foreignaﬀairs.org/articles/iran/2013-06-16/why-rouhani-won-and-why-khamenei-let-

1

notably before he had undertaken any tangible policy decisions–suggests that foreign policy elites
pay close attention to the prospects of shifting interests within regimes. Indeed, the historic opening
of diplomatic relations between the U.S. and Iran–even for only 15 minutes–coupled with a historic
international agreement on Iran’s nuclear program only two years after Rouhani’s election, does
suggest some tentative evidence that a shift–however small–may be taking place.8

While at the time, it was impossible to say whether or not Rouhani was, in fact, a true moderate,
it should be of interest to scholars of international relations that a shift in rhetoric relative to a less
moderate leader prompted such a diplomatic reaction. The shifting sands of international politics
are important to understanding international relations, be this in the ﬁeld of forecasting escalation
and de-escalation of international rivalries (Colaresi, 2005; Colaresi et al., 2008); explaining the
long-term balance of power in the international system (Waltz, 1979; Schweller, 1996; Mearsheimer,
2001); or understanding changes in regime types over time (Levitsky & Way, 2010; Boix, 2011).
These shifts in underlying interests (and the expression thereof) also have direct implications for the
propensity for international attempts at regime change, be this during civil wars (Lemke & Regan,
2004) or as wars of aggression (Crescenzi, 2007), driven by the anticipation of how a given regime
will act on its underlying interests (Singer, 1963; Gent, 2008).

A spate of previous literature has explored various methods of communicating or learning these
interests, from alliances (Morrow, 1986) and voting behavior at the UN (Crescenzi, 2007; Gent,
2007) as signals of foreign policy alignment, to cheaper signals of conﬂict and cooperation (Thyne,
2009). While the former literature has successfully explained a host of outcomes ranging from
conﬂict to civil wars–suggesting that actors do in fact learn from these behaviors–behaviors such as
alliances, votes, and wars, are far removed from the underlying interests in question. For example,
while voting in the UN may be useful for learning about how states relate, these votes are often
marred by politics which distort the very interests that states care to learn about (e.g. through vote
buying; see (Vreeland & Dreher, 2014)). As (Gartzke & Jo, 2006) point out, while students and
8Also of note is the strong domestic pressure Rouhani faces from hard-liners, including from for-
mer President Ahmadinejad himself (http://ifpnews.com/news/politics/nuclear/former-president-
ahmadinejad-reacts-to-iran-nuclear-deal/).

2

professors may actually prefer to drive a Ferrari, their observed actions will likely distort these
preferences since they may only be able to aﬀord a Volkswagen. Similarly, signing alliances and
ﬁghting wars (or not ﬁghting a war–see Fearon (1995)) are almost by deﬁnition distortions of
underlying interests, since they are driven by a large set of constraints that minimizes how these
interests actually manifest.

The latter approach to signaling alignment helps resolve some of these problems. Thyne
(2009), for example, argues that “cheap” signals such as statements of support or oﬀers of foreign
aid can be informative signals of foreign policy alignment. This link between “cheap” signals
and communicating interests is also present in the literature on diplomacy (Ramsay, 2011) and
in economics (Farrell & Gibbons, 1989a,b), though its applications to Political Science have
remained somewhat limited.
In particular, even Thyne’s (2009) approach is limited to direct
signals–for example, President Bush’s speech in 2005 expressing that “America stands with [the
Iranian opposition]”–of opposition and support.9 There are other types of speech that do not speak
as clearly about a speciﬁc side of an issue–such as those at the UN–which may simply make note of
a concerning issue, but are often not as explicit in terms of stating support for one side or another
of an issue.10 These are the types of speeches I analyze in this dissertation.

1.1.1 The United Nations as a domain of analysis for speeches

Politicians speak in a variety of diﬀerent forums. From local media to transnational groups, leaders
express their opinions on a large variety of issues. European Allies, for example, expressed concern
over President George Bush’s use of “Religious language” to describe what he called an “axis of
9His approach relies on a merging of COPDAB and WEIS, two events-data-based measures of

conﬂict and cooperation, which include these “cheaper” signals.

10For example, in his 2015 speech during General Debate, during the 70th session of the UN
General Assembly, President Barack Obama spoke in speciﬁc terms about United States inability
to "stand by while the sovereignty and territorial integrity of [Ukraine] [was] ﬂagrantly violated",
while he spoke in much more general terms about the South China sea, where he proclaimed that
the United States "[makes] no claim on territory there" and that the US "[does not] adjudicate
claims", simply "encouraging China and other claimants to resolve their diﬀerences peacefully"
(UN General Assembly, Session 70 Plenary meeting 13. See UN Document A/70/PV.13, page 10).

3

evil” during his 2002 State of the Union address.11 Conservative hardliners in Iran lambasted the
moderate Rouhani after Trump’s 2017 speech in Saudi Arabia for [Rouhani’s] controversial attempt
at detente with the U.S.,12 and in the same source we note that President Bush’s speech against
Iran helped undermine a moderate Khatami’s “policy of detente. . . [paving] the way to hardliner
Mahmoud Ahmadinejad. . . in 2005.” In other words, when a politician talks, interested audiences
listen–and react accordingly.

A key problem with communicating, however, is comparability: how would we compare two
leaders’ speeches to two entirely diﬀerent media outlets? Leaders speak in both idiosyncratic
terminology with local audiences (for example, using religious language) or they speak speciﬁcally
to their speciﬁc audiences–undermining the ability to compare two speeches. Moreover, the
audiences are diﬀerent, and so there are fewer sacriﬁces to be made: the relative information is
limited in that we are not able to compare how the two politicians would speak in the same setting,
given the same audience. In order to derive useful inferences from speech, we need to be able to
compare leaders in a context with a relatively similar lexicon, audience, and set of issues under
discussion. This is not to say that no information can be gleaned from other sources, but rather it
is to say that a structured context would be ideal.

This problem is analogous to comparing political parties in diﬀerent countries. For example,
how might we determine whether the Democratic Party in the US is more or less liberal than the
Labor party in the UK? To measure this via their speech, we would have to isolate the two parties
speaking in a similar context on similar issues, such that we can gauge how similar or diﬀerent
they are. This would also pose an identiﬁcation problem for deﬁning “liberal”: thus, we might, for
example, compare both to the Republican Party in the United States to determine who is further
away from them on a given issue. In other words, there are two problems to solve when trying to
glean systematic information from speeches: a comparability problem as deﬁned above (i.e. the
inﬂuence of diﬀerent contexts, diﬀerent audiences, etc), and an identiﬁcation problem of estimating

11http://www.pbs.org/wgbh/pages/frontline/shows/blair/etc/cron.html
12http://time.com/4787856/trump-islam-speech-iran-rouhani-conservative/

4

positions on a relative scale (e.g. deﬁning how “liberal” a party is).

This is where institutions can play a role. Institutions coerce leaders to essentially speak to the
same audiences. They also place leaders in the same context, with the same historical context: two
leaders speaking at the United Nations, for example, are drawing from the same UN Charter and
body of international law when crafting their speeches. They also draw from the same history–for
example, the UN’s history with promoting decolonization or, as a speciﬁc example, in founding
the state of Israel (a notably controversial issue at the UN). Moreover, the broad membership of the
UN provides a unique population for comparability. Regional organizations like the EU, AU, etc.
can be valuable–and will likely contain more regional-speciﬁc information–but they do so at the
sacriﬁce of cross-national comparability (for example, comparing France’s speech at the EU with
Liberia’s speech at the AU is less useful than comparing their speech at the UN, even though the
ﬁrst and second are more likely to address more speciﬁc, regional issues).13 Future work would do
well to zoom into these bodies and explore the politics of public deliberation there–with a particular
eye to the audience to whom these bodies are speaking. On the topic of audiences, another unique
opportunity with the UN is its broad political reach: the UN has been one of the preeminent (if not
the preeminent) IGOs of the post-war era, politically relevant enough to reliably reach a massive
audience across the globe. It is a reasonable assumption, then that a very wide, diverse array of
actors–both domestic and international–are at least potentially listening. More importantly, when
two leaders speak publicly at the UN, the same groups should be able to listen to both, creating a
shared political foundation from which to compare the two.

Tying to the literature on diplomacy, formal institutions like the UN are also unique in that
they provide opportunities for both public and private communication. This would maximize
the politically oriented signal from public speeches, and minimize negotiation oriented signals,
since these negotiations often take place in private. For example, in the leadup to the Iraq War
in 2003, the French were able to publicly focus on their disagreement with the United States and
United Kingdom on the Iraq War, while in private they were able to speak in more detail about the
13This links directly to Thompson’s (2009) conceptualization of how a diverse set of interests

can lead to credible signaling at the UN.

5

parameters of this disagreement–for instance, whether Resolution 1441 constituted an “automatic
trigger” for war Recchia (2015, p.634).14 Moreover, the political nature of speech in this case
is illustrated by the united front that France and Germany put up against the United States and
United Kingdom. While the French expressed marginally more support for a possible war than the
Germans–who were categorically opposed to force, even in private15–the two countries put up a
united front in their respective public media pushes and corresponding public speeches16, further
underscoring the way political incentives can be ampliﬁed in public forums where private forums
are equally accessible.

Finally, given potential constraints on the agenda at the UN Security Council (Binder & Golub,
2020), I focus on speeches delivered at the UN General Assembly. This focus also provides the
additional beneﬁt of linking to an ongoing literature on the United Nations, where the bulk of
work has focused on the actions of the UN Security Council (Fortna, 2008; Beardsley, 2013;
Allen & Yuen, 2020). While recent work has begun to open the black box of the UN General
Assembly–especially through the eﬀorts of UN General Debate corpus data collection by Baturo
et al. (2017)–these works often are either descriptive in nature (Baturo & Dasandi, 2017), or assume
the measurement value of these speeches (Chelotti et al., 2018) toward a diﬀerent end. However,
little work exists to understand the speeches themselves, and the political origins of variation in
their magnitude (i.e. how much leaders speak at all) or content.17 In this dissertation, I set out to
lay groundwork toward understanding these origins.

To this end, I analyze three characteristics of speeches at the United Nations General Assem-
bly, each containing potentially distinct but valuable information about the political alignments,
priorities, and interests that countries seek to convey to interested public audiences. In the ﬁrst

14Also see http://www.pbs.org/wgbh/pages/frontline/shows/blair/etc/cron.html
15http://www.spiegel.de/international/germany/berlin-eﬀorts-to-prevent-iraq-invasion-

classiﬁed-papers-prove-german-warnings-to-bush-a-730979.html

16https://www.theguardian.com/world/2003/jan/22/germany.france
17To be explicit, we can think of this as diﬀerentiating between speech as a phenomenon to
be decomposed, speech as an explanatory variable toward an inferential end, and speech as a
dependent variable to be modeled as a function of other political variables. My approach falls in
the third category.

6

essay, I analyze the volume of speech: that is, how much a country speaks at all at the UN General
Assembly. This will provide a broad view of the material and political correlates to the breadth of
issues and depth of opinion that countries express on the array of topics under discussion at the UN
General Assembly. It also provides an insight into whether countries view the United Nations as
a politically valuable forum at all, since their private options would allow them negotiation power
regardless of whether they also utilize the public forum for public-facing political gain.

In the second essay, I analyze the composition of speech: that is, the diversity of issues and
topics that countries speak on at the UN General Assembly. This provides underpinnings to the
political and material incentives from the ﬁrst chapter, and focus on the decision to actually speak
on an issue at all. Moreover, it provides insight into why countries speak on an issue at all; in this
case, I show that the public decision to speak on an issue speaks to that particular issue’s political
importance; this should result in predictable patterns of behavior between countries, since it points
to the presence of shared interests between any two countries (whether they agree or disagree on
that issue).

Finally, in the third essay, I analyze the alignment of speech: that is, patterns of how countries
express their similarities and diﬀerences with other countries across issues.
In any speech, a
country can decide to either reinforce the alignment expressed by other behavior (for instance,
voting against a country while also speaking diﬀerently to them), or they can hedge against that
alignment (for example, voting against a country but using speech to express similarity to them, in
spite of the vote). I show that political conditions surrounding the political costliness of aligning
with a known friend or enemy is correlated with the propensity to use speech in this reinforcing or
hedging manner.

To do this, I collect an original dataset of all speeches delivered at the UN General Assembly
during plenary meetings. This dataset is collected independently of the one collected by Baturo
et al. (2017), and builds upon it by including speeches beyond the General Debate. I detail the data
collection and cleaning process in the next section.

7

1.2 Data collection

The data collection process involved four steps: Scraping the transcripts from the United
Nations repository; Converting the transcripts into machine-readable text; Parsing and cleaning
the transcripts into individual speeches; and Identifying the individual speakers for each speech.

1.2.1 Step 1: Scraping and converting to text

I begin by scraping Portable Document Format (PDF) ﬁles from the United Nations (UN) website.
More speciﬁcally, I access the UN’s oﬃcial document system (ODS), using a URL pattern

http :// www.un.org/en/ga/ search /view_doc.asp? symbol =

followed by the United Nations Document Symbol for the General Assembly transcript, which
is composed of the preﬁx  followed by a session number and a plenary identiﬁer  ., separated
by a slash /. For example, to signify UN General Assembly session 50, Plenary session 99 I use
the URL pattern18

https :// www.un.org/en/ga/ search /view_doc.asp? symbol =A/50/ PV

.99

Using this framework, I wrote a script using Python 2.7 to loop across Sessions and Plenary
Meetings for all sessions starting with session 31 through 69 (i.e. January 1975 through December
2014). I manually identiﬁed the maximum number of Plenary Meetings per session, given in a list
format19

maxlist =[109 , 111, 108, 120, 114, 111, 122, 106, 108, 134, 103,

116, 96, 100, 82, 92, 112, 106, 108, 128, 107, 92, 107, 100,

18This is only applicable to documents starting in 1976. Prior to 1976, the Document Symbol
for General Assembly transcripts would be / . with a single number ranging from 1 through
2444. That is, the earliest Plenary session is given by the Document Symbol / .1 and the last
Plenary session of 1975 (December 17th, 1975) is given by the Document Symbol / .2444.
19I end data collection at /69/ .74. Note that the plenary meetings for 2014 actually end at
/69/ .77

8

112, 112, 94, 95, 118, 101, 109, 122, 105, 122, 118, 131, 99,

109, 74]

Finally, using this URL pattern, I navigate to the page using the BeautifulSoup library in Python
(Richardson, 2007) to set cookies, access the site frame, and navigate to the source location of the
PDF ﬁle, to eventually download the ﬁle. I deﬁne a function to scrape a given URL as deﬁned
above, and download to a ﬁle as deﬁned by the session and plenary meeting. Figure 1.1 provides a
screenshot of the frame in the HTML for a PDF ﬁle, which this scraper is able to access.20

def unga_ scrape (URL , filename ):

BASE_URL = ’http :// www.un.org/en/ga/ search /’

BASE_ ACCESS _URL = ’http :// daccess -ods.un.org ’

# start session

session = requests . Session ()

response = session .get(URL , headers ={’User -Agent ’: ’Mozilla

/5.0 ( Macintosh ; Intel Mac OS X 10_9_4) AppleWebKit /537.36

(KHTML , like Gecko) Chrome /35.0.1916.153 Safari /537.36 ’})

# get frame links

soup = BeautifulSoup ( response .text)

frames = soup.find_all(’frame ’)

header _link , document _link = [ urljoin (BASE_URL , frame.get(’

src ’)) for frame in frames ]

# get header

20Credit to user alecxe on Stackoverﬂow for solving the issue of extracting the PDF from the
UN website’s frame. See https://stackoverﬂow.com/questions/24841632/how-to-get-contents-of-
frames-automatically-if-browser-does-not-support-frames

9

session .get( header _link , headers ={’ Referer ’: URL })

# get document html url

response = session .get( document _link , headers ={’ Referer ’: URL

})

soup = BeautifulSoup ( response .text)

tempmatch = soup.find(’meta ’, content =re. compile (’URL =’))

content = tempmatch [’content ’]

document _html_link = re. search (’URL =(.*) ’, content ).group (1)

document _html_link = urljoin (BASE_ ACCESS _URL , document _html_

link)

# follow html link and get the pdf link

response = session .get( document _html_link)

soup = BeautifulSoup ( response .text)

# get the real document link

content = soup.find(’meta ’, content =re. compile (’URL =’))[’

content ’]

document _link = re. search (’URL =(.*) ’, content ).group (1)

document _link = urljoin (BASE_ ACCESS _URL , document _link)

print document _link

# follow the frame link with login and password first - would

set the important cookie

auth_link = soup.find(’frame ’, {’name ’: ’footer ’})[’src ’]

10

Figure 1.1: HTML for PDF ﬁle on UN website

session .get(auth_link)

# download file (Name example : A/31/ PV.1 = " document 31_1")

with open(filename , ’wb ’) as handle :

response = session .get( document _link , stream =True)

for block in response .iter_ content (1024) :

if not block:

break

handle .write(block)

The only diﬃculty with this method is that the UN website, periodically, kicks oﬀ the scraper.
Thus, I add a unit test to the script to check if a ﬁle exists, and then re-execute the script for each
time it is kicked oﬀ.

1.2.1.1 Conversion to text

To convert PDF ﬁles to text, I use a diﬀerent method to convert ﬁles prior to session 48. Session
48 saw a marked change in the quality of transcript, from previously being typeset as a typewriter
(see Figure 1.2) to being typeset in a more modern two-column layout, with advanced formatting
(see Figure 1.3). For the latter group, I convert the PDF ﬁles to machine-readable text using Adobe
Acrobat Pro, drawing on the Action Wizard to batch convert several ﬁles at a time in a given
directory.

11

Figure 1.2: Pre-Session-48 formatting of PDF transcript

Figure 1.3: Post-Session-47 formatting of PDF transcript

For the former group, I turn to OCR methods.

I use the ABBYY FineReader PDF Optical
Character Recognition (OCR) engine to convert PDF ﬁles between sessions 31 and 47 (ABBYY
Production LLC, 2013). To convert these at scale, I use a Mac OSX Automator to set up a folder
action, implementing the ABBYY FineReader conversion to text for any PDF ﬁle that exists in that
folder (see Figure 1.4). Next, I use a BASH script to automate the process of transferring individual
PDF ﬁles to the folder in question, sleeping for several seconds in between each to allow the text
conversion to take place, and deleting the PDF ﬁle from the folder once the text ﬁle appears in the
output folder. I repeat this process until all PDF ﬁles have been converted to machine readable

12

Figure 1.4: Mac Folder Action for ABBYY FineReader

Figure 1.5: Machine readable text ﬁle with speeches highlighted

text.21

1.2.2 Step 2: Parsing and Cleaning

The bulk of the work to collect UN General Assembly speeches occurred in this stage. The
machine readable text for both pre-session-48 and post-session-47 contains a continuous stream of
text throughout the transcript. Recall, the goal here is to compile and identify individual, unique
speeches for every transcript. Figure 1.5 shows an example of what this looks like with respect to
the machine-readable text ﬁles, highlighting the individual speeches visible in the snapshot.

21A brief comparison of Adobe Acrobat Reader’s OCR engine and ABBYY’s OCR engine
revealed that the latter seemed to perform better, especially given the formatting issues in the
pre-48 group.

13

Despite the fact that these texts are continuous streams of text, there is a key characteristic
that makes parsing possible: every speech begins on a new line, and these new line characters
are preserved in the conversion from PDF to Text. For example, as Figure 1.5 shows, despite the
fact that the ﬁrst speech is multiple lines long, the next speech begins on a new line.22 This new
line character serves as the ﬁrst pillar of parsing the continuous text into individual speech-level
observations.23

Second, notice from Figure 1.5 that the speaker–after the new line–tends to end with a colon
(" "). This is only a tendency, though: the second speech in this snapshot shows how, especially
in pre-session-48 observations, but generally across all transcripts–there are times when the PDF-
to-Text conversion misreads a colon as another symbol, such as an exclamation point "!" or a
semi-colon ";". Moreover, there are other idiosyncrasies that exist in some documents, such as
rogue symbols being falsely detected immediately after the new line and prior to the ﬁrst letter of
the speaker name; new lines in the middle of a speaker’s name; and inconsistencies in the detection
of whitespace and tabs between words. To this end, I deﬁne a regular expressions function as
deﬁned below, using the regex library in Python24

def speechparse (sp_text): #sp_text = speechtext 1b

’’’This function takes continuous stream of UNGA text and

returns

22Readers will also note that the ﬁrst speech contains non-speech, administrative text about a
speaker being escorted into the hall. As discussed later, however, this is presidential speech–which
is usually removed from analysis since the function of their speech is by deﬁnition administrative
as opposed to as a representative of their countries. Future work may revisit these presidential
speeches to determine whether systematic diﬀerences in content exist depending on the country
holding the presidency.

23To be thorough, I also clean the entire text ﬁle prior to parsing the individual speeches. This
includes substituting various OCR mistakes that were manually detected during the process of data
cleaning, such as replacing "iVIr" with the correct "Mr.", or replacing excess whitespaces, tabs, or
new lines with simple spaces. It also replaces speciﬁc instances where symbols would interfere
with the parser, such as replacing apostrophes in known phrases (so that they are not picked up by
the speech parser), etc. This script is available upon request; it consists of approximately 60 lines
of regular expressions substitution statements.

24This script was written in Python 3.

14

a list of individual speeches per the regex rules below.’’’

speechtext 3=re.sub(r ’.*? The meeting was called to order ’,’’,

sp_text ,flags=re. DOTALL |re. MULTILINE |re. IGNORECASE )

speech =re. findall (’\n +[\ -\}\{\|\ s\t\r\n]*([A-Z][^\/\ ,\ ’\n\r

]+?[\n\r]{0 ,4}?[^\/\ ,\ ’\n\r ]+[\:\;\!]) (?=\s{0 ,2}[A-Za -z

12580]) (.*?) (?=\n +[\ -\}\{\|\ s\t\r\n]*[A-Z][^\/\ ,\ ’\n\r

]+?[\n\r]{0 ,4}?[^\/\ ,\ ’\n\r ]+[\:\;\!]\ s{0 ,2}[A-Za -z

12580]|\ Z)’,speechtext 3,re. MULTILINE |re. DOTALL )

if len( speech )==0:

speech =None

return ( speech )

Each transcript also contains meta-data about the meeting time–i.e.

the date of the meeting,
among other meta-data such as the time of day the deliberations began and ended. I also parse these
using regular expressions, where I compile a manual list of plenary sessions which had parsing
errors. This is a tactic which I also used during parsing of speeches in the previous step: I used
the more speciﬁc functions on all data to isolate the speciﬁc information desired (such as the date
of the meeting), and after manual validation, I use a more general function to at least capture this
information–even if it means sometimes also capturing more noise. This iterative method provides
some balance between ﬂexibility of capturing more data and the speciﬁcity of actually isolating the
data of interest.

An important characteristic of this parser, is that it splits speeches by looking for speakers in
the text. Thus, the parser will return both speakers and speeches simultaneously. This will become
important in the cleaning step, when falsely split speeches will be identiﬁed by utilizing the content
of the speaker.

if str(ses)+"_"+ str(pv) in parse_errs:

15

meettimetxt =re. findall (’\A.*? General [\s\n]+ Assembly

.*?([A-Z][^\n]+\n?[^\n]+m\.\s).*? President \s*’,

text ,re. DOTALL |re. MULTILINE |re. IGNORECASE )

meettimetxt 2=re. findall (’[ MTWFS ][a-z]+ day [^\n]+\n?[^\

n]+m\.’,text ,re. DOTALL |re. MULTILINE )

else:

meettimetxt =re. findall (’\A.*? General [\s\n]+ Assembly

.*? Session .*? Meeting .*?([A-Z][^\n]+\n?[^\n]+m\.\s)

.*? President \s [\:\;\ I]*’,text ,re. DOTALL |re.

MULTILINE |re. IGNORECASE )

meettimetxt 2=re. findall (’[ MTWFS ][a-z]+ day [^\n]+\n?[^\

n]+m\.’,text ,re. DOTALL |re. MULTILINE )

if meettimetxt :

meettime = meettimetxt

elif meettimetxt 2:

meettime = meettimetxt 2

else:

meeterr =’ Meeting time not found ’

meettime =[]

errs. append ( meeterr )

This parsing method outputs a Comma-Separated-Values (CSV) ﬁle for individual plenary
sessions, containing unique observations (i.e. rows) for each split identiﬁed by the parsing function
deﬁned above.

16

1.2.2.1 Cleaning: parsing error correction

The parser as deﬁned above provides a useful start for isolating individual speeches. However, it
is not suﬃcient. Any parsing will require a balance between (a) correctly capturing all the true
splits (that is, making sure two diﬀerent speeches are not labeled as being the same speech); and (b)
preventing any false splits (that is, making sure a single speech is not falsely labeled as being two
diﬀerent speeches). In the parser above, I favored the ﬁrst goal (a) more heavily than the second,
with the intention of following up in later steps to rectify the erroneous splits in (b).

To identify erroneous splits, I take advantage of the ordinal nature of the parser, and the fact
that we have both speaker and speech information. For example, in Figure 1.5, suppose the second
line of the last speech (blue) is falsely identiﬁed as a unique speech, where the phrase "presidency
of the forty-seventh session of the General Assembly" is falsely identiﬁed as a new speaker, and
the text afterward identiﬁed as the corresponding speech. Given that we can identify this speaker
as not actually being a speaker, we can concatenate both this "speaker" and "speech" with the
previous speech (which in this example, assume is correctly identiﬁed with the speaker "Chairman
GORBUNOVS (spoke in Latvian; English text furnished by the delegation)"), resulting in a single
speech containing what was previously the correctly identiﬁed speaker, the correct corresponding
speech, a falsely identiﬁed speaker, and a falsely corresponding speech in that order.

Identifying these erroneous speakers was not a simple task, and required a manual reading
of each identiﬁed speaker to determine whether it is legitimate or illegitimate. To facilitate this
manual reading, I use a three-step process to identify thresholds above and below which we can
separate legitimate speakers from illegitimate speakers. More precisely:

1. In step 1, I ﬁnd the cutoﬀ number of words (for example, 5 words) above which I only

capture non-speakers, looking for for PERFECT SEPARATION in that cutoﬀ.

2. In step 2, I ﬁnd the cutoﬀ number of words (for example, 4 words) below which I captured
ALL legitimate speakers. This process includes manually compiling an exception list for

17

legitimate speakers, where–for example–if we have one or two speakers above the identiﬁed
cutoﬀ, I can include them as exceptions so they’ll be included in the ﬁnal list.

3. For step 3, sometimes there are a lot of speakers above a given cutoﬀ. For example, above
3 words we still get a ton of speakers, but you’ll also get a lot of non-speakers. This creates
diﬃculty using an "exceptions" list. Therefore, this step looks for a cutoﬀ below which I
capture all legitimate speakers, where we specify an exception list for illegitimate speakers
to remove from below that cutoﬀ.

To do this, I ﬁrst stack all individual plenary meeting CSV ﬁles into a single session-level CSV
ﬁle. Then, using trial and error, for each session I execute the three steps above. For example, for
step 1 on UNGA session 37, I will compile all speakers in that session and manually review all
speakers above 3 words long. If there are any legitimate speakers in that list, I increase this to 7
words. If there are no legitimate speakers in that list, I will decrease it to 5 words. If there are still
no legitimate speakers, I will decrease to 4 words. Finally, if there are legitimate speakers present
above the 4-word threshold, I will set the threshold for session 37 for step 1 as being 5 words. For
example, a resulting source ﬁle for thresholds would look like this in Python list format, for all
sessions from session 37 through 69:

[5, 12, 10, 21, 19, 14, 15, 18, 16, 15, 18, 6, 9, 7, 3, 3, 3, 4,

4, 3, 3, 4, 4, 6, 5, 4, 4, 5, 5, 9, 9, 4, 5]

I then repeat this process for each step, where step 2 and step 3 also include manually compiling
exceptions to the thresholds–as deﬁned in the above summary. More precisely, this list of exceptions
would be the exact text of speakers that are identiﬁed to be exceptions to thresholds in step 2 and 3.
Finally, after all source ﬁles (i.e. thresholds and exceptions lists) have been identiﬁed, I run a python
script for each iteration, where erroneous speakers–and the corresponding erroneous speech–are
concatenated with the previous speaker/speech observations. Note that I do this as a loop: this
means that if two consecutive rows are identiﬁed as erroneous speaker/speech observations, the

18

Figure 1.6: Example speakers with parenthetical identiﬁcation

ﬁrst of these errors will be concatenated with the correct observation before it, and then the second
of the original errors would be concatenated with that newly concatenated, speech.

Ultimately, while no parsing methodology will be exact, this methodology provides a useful
balance between automated identiﬁcation and parsing with the precision of manual reading. This
marriage between human and machine is an important part of the data collection for this dissertation,
and is a conceptual framework for data collection that should be replicated in other areas of large-
scale data collection and cleaning.

1.2.3 Step 3: Identifying speeches with individual countries

Having compiled speech-level observations following the methodology above, I turn to the task of
labeling each speech with a respective speaker. Many speakers contain this information within their
text; as Figure 1.6 shows, identiﬁcations of speakers with countries are often listed parenthetically
after the speaker name and prior to the colon separating speaker and speech. For these, it is
relatively straightforward to use a Regular Expression to search the speaker for parentheses and
search for matches between the text within parentheses and a source ﬁle linking the text of country
names to Correlates of War (COW) country codes.

A large number of speeches, however, do not contain this parenthetical information. For
example, Figure 1.7 United States representative to the UN Madeleine Albright’s speech is not
tagged with a corresponding parenthetical label. However, the opportunity to identify these speeches

19

Figure 1.7: Example speaker without identiﬁcation

Figure 1.8: CSV ﬁle rows with speakers and speeches

comes from the speech immediately before: I exploit the institutional formalities of the UN General
Assembly to use the immediately preceding speech to identify unidentiﬁed speeches. In nearly
every case, this preceding speech is the UN General Assembly president introducing the speaker–
including their respective country or organization.

To this end, I loop sequentially over every speech in my cleaned dataset.

I use a regular
expression to attempt a match on whether or not the speaker name contains parentheses, and then
I attempt to match the text within those parentheses with country names in a source ﬁle mapping
If there is no match, I pull both that speaker-speech
country names to COW country codes.
observation and I pull the speaker/speech observation immediately preceding it. This resulted
in a CSV ﬁle with the text of both speeches one after another (Figure 1.8), where the country
or organization name can be manually added.25
I then add country names manually to each
unidentiﬁed speaker-speech observation and save the CSV ﬁle as a source ﬁle; then, I re-run the
sequential loop over all speeches, drawing from the source ﬁle for those speakers which do not
contain country names in parentheses.

Finally, a critically important phenomenon that occurs in UN General Assembly speeches are
grouped statements. These statements are delivered by one country on behalf of a particular group–
for example, on behalf of the European Union. This becomes problematic for country-level analyses
since it censors the content of speech in transcripts: for example, if Canada delivers a speech on

25Thanks to Tara Iseneker for assistance with this

20

behalf of Canada, Australia, and New Zealand, then the data would only show this speech as
being associated with Canada. This censors the fact that, by deﬁnition, all three countries–Canada,
Australia, and New Zealand–eﬀectively delivered the same speech in that instant. This would,
for example, systematically underestimate the similarity of speeches between two countries who
deliver speeches via the same grouped statement, and impact the subsequent analysis linking speech
patterns to foreign policy or political incentives.

To this end, I also exploit the formal nature of speeches at the UN General Assembly by
searching for common phrases at the start of a speech, indicating it is a grouped statement. To do
this, I search for the phrases "speak on behalf of" or "in my capacity as", both of which almost
always precede grouped statements.26 I also compile source ﬁles for every group I ﬁnd in the data;
Table 1.1 provides a list of all of these. Notice that many are draft resolutions: this is due to the
fact that in many cases, a country will deliver a speech on behalf of the signatories to a particular
draft resolution. Similarly, there are groups such as African Parties to the International Criminal
Court (ICC), which are exactly as the name describes–African members of the ICC.

26While I cannot claim to have read every speech to validate this tactic, there were no grouped

statements I came across which did not include these phrases.

21

Latin American and Carribean Group CentAmerica
MERCOSUR
NAM
Nordic
OIC
PIF
PSIDS
Pompidou

CentralAmericanIntegrationSystem
Committee on the inalienable rights of the Palestinian people
Democracy and Human Rights in Haiti
Draft Resolution A51L71
Draft Resolution A52L46Rev1
Draft Resolution A52L67Rev1
Draft Resolution A53L16
Draft Resolution A59L54
Draft Resolution A68L33
Entrepreneurship and privatization for economic growth and sustainable development
FriendsOfMediation

ALBA
AOSIS
ASEAN
AU
AfPartToICC
AfricanGroup
AntarcticTreaty
ArabLeague
ArabMaghrebUnion PortugueseSpeaking
AsiaGroup
CARICOM
CIS
CPLP
CSCE
CSTO
ECCAS
EEC
EEG
EU
G77
G77andNAM
GUUAM
Group77china
IberoAmerican
LDCs
LLDCs

RioGroup
SACM
SADC
SIDS
SOPAC
ShangCoop
SpanishSpeaking
Union of South American Nations
WEOG
CANZ

Table 1.1: Groups found in UN General Assembly speeches

22

As the python script loops over each speech, I include a prompt to manually enter a source ﬁle
for a given speech, given that it is caught by the phrases above. I then read the caught sentences
and manually input a corresponding source ﬁle: for example, if a speech is delivered on behalf of
the European Union, I input "EU.csv" to indicate that the speech is referencing the list of countries
given by the source ﬁle EU.csv. Further, in many cases, additional countries will also sign on to
a statement–even if they are not members of the group.
I also list these countries manually as
prompted in the Python script. Depending on the organization, I also include the year–and in some
cases, the months–that individual countries actually became members of the group. I add this into
the source ﬁles, and write a function accordingly to read the CSV ﬁle correctly depending on the
group. This returns a list of countries–both within the organization, and any additional countries
who signed on to the speech–which I then use to repeat the speech in question and create separate
country-speech observations for every country in the list.

def csv_ident_ extraction (fname ,sourcedir ,yr ,mo ,type1,type 2):

_ curdir =os. getcwd ()

os.chdir( sourcedir )

_ outcntrylist =[]

_app =[]

if ";" in fname:

_z=fname.split (";")

fname=str (_z[0])

_app =_z[1:]

#Cross ref with file types

if fname not in type 1+ type 2:

_tp=3

elif fname in type 1:

23

_tp=1

elif fname in type 2:

_tp=2

with open(fname ,’r’) as _f:

_ reader =csv. reader (_f)

for _row in _ reader :

_cntry =_ row [0]

if int (_tp)==1:

_ joinyr =int (_ row [2])

_ joinmo =int (_ row [1])

if int(yr) >=_ joinyr and int(mo) >=_ joinmo :

_ outcntrylist . append (_ cntry)

elif int (_tp ==2):

_ joinyr =int (_ row [1])

_endyr=int (_ row [2])

if int(yr) >=_ joinyr & int(yr) <=_ endyr:

_ outcntrylist . append (_ cntry)

elif int (_tp ==3):

_ joinyr =int (_ row [1])

if int(yr) >=_ joinyr :

_ outcntrylist . append (_ cntry)

else:

sys.exit (" csv_ident_ extraction () function

requires type 1;2; or 3")

_ outcntrylist . extend (_ app)

os.chdir (_ curdir )

return _ outcntrylist

24

if " speak on behalf of" in first.lower () or "in my capacity as"

in first.lower ():

cap=re. search (" speak\s+on\s+ behalf \s+of\s*(.*)",first.lower ())

try:

ident =( int(ses),int(pv),str(spk),str(cap.group (1)))

except AttributeError :

cap=re. search ("in\s+my\s+ capacity \s+as\s*(.*)",first.

lower ())

ident =( int(ses),int(pv),str(spk),str(cap.group (1)))

print(ident)

challenge =input (" file ?_")

challenge =comma2 semicolon ( challenge )

if challenge =="":

challenge =input (" Are you sure ?")

if challenge =="":

#We still need to put it in our dict to confirm

assert grouped ==0

challenge =ccode

groupdict [ident ]= challenge

newlab =ccode

finrow =[ses ,pv]+ rowdata +[ newlab ,grouped ,corrccode ,

speech ]

writerow (finrow ,outfile ,outdir ,’a’)

counter +=1

continue

25

Figure 1.9: Total word count over time

else:

pass

#if I give it a value for the input , go find it

grouped = challenge

groupdict [ident ]= challenge

ungrouped = GetUngrouping (challenge ,sourcedir ,yr ,mo ,type1,type 2)

for cntry in ungrouped :

newlab =cntry

finrow =[ses ,pv]+ rowdata +[ newlab ,grouped ,corrccode , speech ]

writerow (finrow ,outfile ,outdir ,’a’)

The data consist of 84,872 speeches delivered by individual countries between 1982 and 2014.27
These consist of over 81,000,000 words spoken, with an average of around 1.2 million words per
year prior to 1992, and an average of around 3 million words from 1992 onward. Figure 1.9 plots
the total yearly word count over time; similar trends are also present for the number of speeches
delivered per year.

27Slightly over 50,000 speeches were delivered by others, including organizations; celebrities;

or other speakers

26

Figure 1.10: Mean word count per speech, over time

One noteworthy pattern is that despite the fact that total words and speeches have increased
over time, the average length of any given speech has decreased. Figure 1.10 plots the average
length of any given speech, per year, over time. Thus, the increase in speech at the UN General
Assembly over time is largely driven by more individual speeches, which is likely driven by more
issues being discussed; this follows from previous work on the UN General Assembly by scholars
studying voting patterns (Bailey et al., 2015), who show that the agenda has changed over time as
countries vote on diﬀerent issues.

27

CHAPTER 2

MATERIAL CAPACITY AND POLITICAL INCENTIVES AS DRIVERS OF UN

GENERAL ASSEMBLY SPEECHES

Since its inception in 1945, the United Nations has captured the attention of both scholars and
policymakers as a vehicle for countries to channel their foreign policy interests. The ensuing
discussion has been a diverse one: from speciﬁc behaviors like peacekeeping and other policies at
the UN (Neack, 1995; Fortna, 2008; Beardsley & Schmidt, 2012), to institutional questions about
how these policies might come about (and their impact on international politics) (Voeten, 2001;
Mearsheimer, 1995; Keohane & Martin, 1995; Mearsheimer, 1994), to the very sources of the
UN’s political legitimacy itself (Keohane, 2012; Chapman, 2011; Thompson, 2009; Ruggie, 1982;
Claude, 1966). This body of work has provided a rich picture of the UN’s role in international
politics, particularly with respect to the way the UN promotes and legitimizes the core values of its
organizational mission.

While some have argued that the institution itself can take on an agenda of its own as an "active
participant" in world aﬀairs (Barnett & Finnemore, 2004, p.16), the UN is composed of individual
actors. The individual member states of the UN engage with the UN in a variety of diverse ways. A
recent literature, for example, has examined the patterns of how individual countries contribute to
United Nations peacekeeping missions (Kathman & Melin, 2017; Bellamy & Williams, 2013; Bove
& Elia, 2011), in relation to broader theories on where the UN sends peacekeepers more generally
(Uzonyi, 2015; Fortna, 2008; Gilligan & Stedman, 2003; Andersson, 2002). This literature has
shown that countries contribute troops where the risk of having idle troops otherwise would be
high: Kathman & Melin (2017), most notably, show that the history of coup attempts as well
as the presence of a rivalry both are strongly related to the number of troops contributed to UN
Peacekeeping missions. Another vehicle to understand individual state engagement with the UN
has been the study of roll call votes at the UN General Assembly (Ball, 1951; Lĳphart, 1963; Alker,
1964; Russett, 1966; Keohane, 1969; Newcombe et al., 1970; Voeten, 2000; Bailey et al., 2015),

28

with some notable work also looking at patterns of resolution sponsorship at the UNGA between
countries (Jacobsen, 1969; Dĳkhuizen & Onderco, 2019). Finally, a literature has arisen studying
the dynamics of voting itself: an argument has emerged, for example, that powerful countries are
able to buy votes from weaker countries at the UN General Assembly Carter & Stone (2015);
Dreher et al. (2008); Wang (1999); Rai (1980). In both the cases of Peacekeeping and General
Assembly voting, current literature has shown that both political interests and material capacity
can help drive and explain individual state engagement with the UN.

Another type of behavior at the UN has yet remained overlooked, despite its proliﬁc nature going
well beyond both voting and Peacekeeping: individual country speeches at the UN–and in particular,
the UN General Assembly. In 2014 alone, there were over 2.2 million words spoken on the ﬂoor
of the UN General Assembly, composed of 194 countries and over 3500 unique speeches. These
speeches occur in a variety of contexts on a plethora of issues, ranging from Economic development
to Nuclear proliferation, and are one of the most politically visible actions an individual country
can take at the UN. Yet to date, in the scholarly literature, little systematic study has been done
on this behavior; while some work has alluded to the presence of a public forum (Voeten, 2005),
and others have made use of the speeches themselves to answer other questions (Baturo et al.,
2017; Medzihorsky et al., 2017; Binder & Heupel, 2015), the actual variation between countries
in their delivery of speeches has not yet been subject to empirical scrutiny. Especially given the
public nature of these speeches–particularly during Plenary sessions–and the increasingly proliﬁc
nature of these speeches in the media and political sphere more broadly,1 this behavior provides an
opportunity to more deeply study the politics of public diplomacy at the United Nations.

In this paper, I set out to examine why some countries speak more at the UN General Assembly
than other countries. I argue that speech at the UN General Assembly is related to the audiences
who may potentially be listening, and that countries with greater capacity to inﬂuence issues
will gain more political beneﬁt in speaking on them (resulting in more speech). First, I show
1A search of New York Times articles for "U.N. General Assembly" since 1990 returns over
7500 results, with an average of 16 per month from 1990-2001, and an average of 24 per month
from 2001 to August 2020

29

that countries with a greater capacity to inﬂuence the UN’s broader institutional mission–namely,
"international peace and security" and the "relations among nations"2–should speak to a more
diverse set of audiences, since they will either be involved in or have the capacity to be involved
in more issues relevant to the UN. Second, I show that countries with stronger political incentives
linking audiences to leaders–in this case, through domestic political institutions–will speak more
than their counterparts, all else being equal. Third, I evaluate the mechanisms linking both material
capacity and political incentives to the amount of speech, showing that powerful countries who
show an explicit interest in a larger number of issues speak the most, and that the impact of political
incentives on the amount of speech is highest for those countries with fewer alternative avenues
for delivering these speeches. Finally, I validate that the amount of speech at the United Nations
General Assembly does indeed reﬂect an interest in a broader array of issues relevant to the UN as
a whole, as well as an interest in inﬂuencing these issues. I do so by drawing from the literature on
UN Peacekeeping troop contributions, showing that countries who speak more are strongly more
likely to both participate in missions and contribute more troops to it. This suggests that the amount
of speech at the UN General Assembly is directly linked to the issues the UN is most engaged with,
and that it is driven politically by an interest in inﬂuencing these issues.

2.1 The UN General Assembly as a platform for connecting speakers to

audiences

There is a fair amount of literature on rhetoric in international politics, not only limited to
the UN General Assembly. This literature is diverse, analyzing international speech through
multiple lenses: from justifying one’s own behavior to an interested political audience (Steﬀek,
2003, drawing from Weber (1978) and Habermas (1984)), to coercing others into a given behavior
(Fearon, 1994; Schultz, 1998; Tomz, 2007), to even persuading others (Krebs & Jackson, 2007;
Steele, 2007). While each of these literatures are robust and ongoing, all three point to a key insight
about rhetoric: namely, that rhetoric does not occur in a vaccum. Rather, a fundamental component
of any analysis of speech must include a link between the speech being delivered and the audiences

2(Nations, 1945, art.1, par. 1)

30

who are listening or potentially listening.3

Through this lens of linking speaker to audience, the politics surrounding UN General Assembly
speeches can be more clearly understood. Both domestic and international audiences have some
stake in leaders’ behavior and interests at the UN broadly–both at the Security Council and the
General Assembly. From seeking legitimization of force at the UN Security Council (Claude, 1966;
Voeten, 2005; Thompson, 2009; Chapman, 2011) to the coercive politics of vote-buying at the UN
General Assembly (Carter & Stone, 2015; Dreher et al., 2008; Wang, 1999; Rai, 1980), there is a
growing body of work pointing to the fact that countries behave at least as if there are audiences
interested in both the costly and symbolic behaviors at the UN.4

UNGA Speeches almost certainly fall into the camp of symbolic behavior. With minimal if
any cost involved in their delivery, UN speeches are entirely symbolic actions.5 Like General
Assembly votes, given the UNGA’s lack of legal power, speeches are almost entirely political
and serve a communicative function; often crafted by speechwriters or leaders themselves, these
speeches provide leaders a mechanism for targeting key political demographics, speaking to media
at home and and abroad, and communicating values on a broad array of issues (Collier, 2018;
Hecht, 2016; Waheed et al., 2013). Here, governments are free from "external constraint", whereby
3Despite its association in the literature with "audience costs", the term "audience" here is used
more broadly, indicating any actor or group of actors who either listen to or can potentially listen
to a speech.

4To be clear, this literature does provide some evidence of audiences actually both listening
and reacting to actions taken by the UN (Chapman, 2011). Moreover, vote buying is by deﬁnition
costly, even to powerful countries like the US: given that UNGA votes do not carry any legal power,
this must mean that there are explicit and tangible beneﬁts (to oﬀset the costs) from symbolic
communication like UNGA voting. These beneﬁts are only possible if politically relevant audiences
(such as donors and elite interest groups) actually internalize symbolic behavior, like UNGA voting
patterns. Thus, there is at least some evidence that states not only behave as if politically relevant
audiences are listening and watching, but that they in fact are listening and watching.

5This is an assumption, but not without meritorious challenges. The role of speechwriters in the
United States, for example, demonstrates the possible moderating inﬂuence of political institutions
on the actual composition of speech. This adds a non-trivial cost to crafting a political speech, since
it requires investment in both educating speechwriters and in the political infrastructure required to
analyze relevant values and political audiences. While it is beyond the scope of this paper, future
work may consider how these costs impact the subsequent political incentives around delivering
speeches.

31

they are empowered to "speak their mind" rather than being constrained by institutional processes
(for example, voting) or coercion (Baturo et al., 2017, p.2, citing Waltz (1959); Smith (2006)).6

Finally, an important characteristic of a public forum like the UN General Assembly is the
scope of issues it addresses, and the corresponding scope of audiences it can reach–each with a set
of one or more issues they prioritize. Both domestic and international audiences are exposed to the
content, character, and volume of speeches that countries choose to deliver. This also diﬀerentiates
speeches delivered publicly at the UNGA from dialogue that occurs in private: public speeches at
the UNGA serve the purpose of communicating eﬃciently to political audiences who otherwise
would not be present in a private setting.7 Thus, while private dialogue and discussions are a
vital function of institutions like the UN, it is not immediately obvious that they can serve as a
replacement for public speech, and so public speech can be understood independently of these
private options.

2.1.1 Material capacity, Political institutions, and UNGA speech

With the lack of constraint involved with speeches, as well as the communicative function of
speeches to audiences, we should expect two things. First, an increase in the number of politically
relevant audiences, each presumably with their own set of prioritized foreign policy issues, should
correspond with an increased volume of speech at the UN General Assembly–holding all else
equal–given that these leaders must address a more diverse array of issues. Second, a stronger
political connection between leaders and politically relevant audiences should also lead to an
increased amount of speech, since leaders will not only need to elaborate on their positions more
clearly, but they will also need to explain their positions further to dissenting audiences.

Given the global scope of the United Nations, in particular its mandate for "international peace
and security" and the "relations among nations"8, a country’s material capacity serves as a natural

6This is analogous to what Gartzke & Jo (2006) argue with respect to UNGA voting.
7This distinction serves as a foundational element in analyzing private diplomacy, both in its

incidence (Baum, 2004) and its eﬀectiveness (Kurizaki, 2007)

8(Nations, 1945, art.1, par. 1)

32

barrier for its capacity to inﬂuence–and thus, to be politically relevant in–the issues most relevant
to the United Nations. As discussed extensively in previous literature, from peacekeeping to even
symbolic votes on UNGA resolutions, the most powerful countries are consistently in a position to
inﬂuence a more diverse array of issues relevant to the UN. Both the capacity to coerce less powerful
countries to bend to their political will as well as the capacity to inﬂuence the issues themselves (for
example, the capacity to pressure nuclear countries to de-nuclearize or limit proliferation) makes
material capacity a useful measure for the diversity of issues in which that country is involved,
interested, or potentially relevant at the UN. It should follow, then, that these countries will speak
the most at the UNGA, since there are a larger number of politically interested audiences (by virtue
of a more diverse set of issues the country is interested in) present for that leader.

Hypothesis 1: More powerful countries will speak more at the UN General Assembly

Additionally, within any given political audience, we would expect some diversity of preferences,
since any single issue will have a diverse possible array of potential responses, dimensions, and
political angles.9 Thus, leaders with institutional incentives leaving them beholden–or politically
vulnerable–to interested audiences should be more active in their speech, since they must provide
suﬃcient elaboration to all sides of an issue.10 It should follow, then, that countries with domestic
institutions connecting leaders to constituents more strongly–such as democratic countries–will
tend to speak more at the UNGA.

Hypothesis 2: More democratic countries will speak more at the UN General Assembly

9For example, a given issue might see some hawkish and some dovish sub-constituents, each

with some level of political power that inﬂuences a leader’s public speech (Kertzer, 2016).

10As a corollary, this also implies that they will discuss a more diverse array of issues, since
they are more beholden to these diverse audiences, which further suggests an increased amount of
speech. See Mahmood (2020a) for more.

33

2.2 Data and measurement

Data on these speeches come from the UN website, where I scrape all transcripts for plenary
sessions in PDF format.11, converting these to text via Adobe Acrobat Pro’s batch OCR processing.
Next, I identify each speech with its corresponding country and speaker using a combination
of regular expressions (where speakers are identiﬁed parenthetically) and manual coding, using
Python to write an interactive script to detect unidentiﬁed speakers and pull the preceding speech–
the vast majority of the time being an introductory speech by the UNGA president. I then manually
identify the speaker from this preceding introductory speech, and re-process it into the full dataset
programmatically, integrating in the manually identiﬁed speakers as source ﬁles. This corpus was
collected independently of the one by Baturo et al. (2017), whose corpus focuses only on speeches
delivered by heads of state under General Debate.12 The time frame for this paper is 1984-2014.13
Given that the UNGA is a ﬁnite medium, there may be some slight right-censorship: that is,
countries who may have otherwise spoken more might constrain themselves. However, these are
not substantial constraints, and should not have a substantial impact on the analysis14. Moreover,
if in fact there are substantial numbers of countries delivering high-volume speeches who would
have otherwise delivered more, this would only diminish the variation between countries making
empirical leverage more diﬃcult to gain. Finally, as Baturo et al. (2017) point out, UNGA speeches
11This is achieved by exploiting the fact that UNGA plenary sessions are labeled consistently in
the format A/[session]/PV.[plenary], across sessions and plenary meetings within each session.
This is conducted in Python, and draws directly from the UN website, www.un.org

12For example, their data from the year 2000 (session 55) includes the speech made by Madeleine
Albright in the 10th plenary meeting of the 55th session, but leaves out President Clinton’s address
during the 3rd plenary meeting, which occurs outside the conﬁnes of "General Debate" and under
diﬀerent agenda topics.

13Speech data were originally collected to correspond with voting data, which only goes back to

1984. Data has been collected for the years 1945-1984, but parsing and cleaning is ongoing.

14Only the General Debate has a "voluntary" limit of 15 minutes on speeches. The broader plenary
sessions do not have an explicit time limit set, though there are rules in place to allow member
states to vote on placing one should they decide to. As I discuss below, my data encompasses
all speeches including both General Debate and Plenary meetings. For more information, see the
Rules of procedure for the General Assembly, focusing primarily on Rule 72 in the Plenary Session
rules. Also see the Frequently Asked Questions on the UN General Assembly.

34

are unique in that "...the lack of external constraints means that when delivering their GD statements,
governments have more leverage with the positions they take and the issues they emphasise" Baturo
et al. (2017, p.2), meaning that individual country word counts have the added empirical beneﬁt
of being relatively independent observations, which facilitate the use of econometric and statistical
tools and the underlying assumptions therein.15

Between 1982 and 2014, a total of around 59 million words were spoken, with an average of
around 1.8 million per year across all countries. In any given year, there are usually around 1200
speeches delivered across all countries at the UNGA, ranging from around 750 as the Cold War
came to a close in 1989 to around 1500 in multiple years as the UN’s activity increased. The
average country delivered around 8000-10000 words per year, though there is substantial variation
across countries. Figure 2.1 provides the total yearly word count across this period.16. The variation
across countries is visualized in Figure 2.2, which provides the average yearly word count per year
across the world from 1982-2014.17

In the aggregate, there appears to be strong support for the theory that both powerful and
Liberal democratic countries are speaking more with at the UN General Assembly: in addition
to the US, Russia, and China, countries in Western Europe are highly engaged with the UN.
Regional powers also play a signiﬁcant role: Brazil in Latin America, Egypt and North Africa more
broadly, South Africa, India and Pakistan, and Indonesia are all relatively high in UN engagement.
Figure 2.4 provides a map of European Union countries in their respective word count; Figure 2.3
and Figure 2.5 provide similar maps for countries in the non-aligned movement and in Africa.

15As I discuss later, a key source of non-independence comes from the presence of grouped

statements, which I account for in the analysis.

16Individual country averages follow a similar trend
17Mapping is done via the R package rworldmaps. See South (2011).

35

Figure 2.1: Total word count, yearly

2.2.1 Measurement of country-level characteristics

2.2.1.1 Variables of interest

To measure material capacity and power, I use two diﬀerent measures, each capturing a diﬀerent
aspect of power. First, I use the Composite Indicator of National Capabilities (CINC) (Singer et al.,
1972, version 5), which provides as a continuous variable an indexed score aggregating military,
population, economic, and other measures of a country’s general level of power. This provides a
broad, comparable measure of power and material capcity across countries. Noteworthy here is to
point out that the component measures of CINC score–especially GDP Per Capita, Population, and
Military Spending–can each be analyzed separately as diﬀerent components of broader "material
capacity". As noted below as well, the high correlation between CINC score and these constituent
components (in particular, population, which has an over 0.90 correlation with CINC score) makes
it diﬃcult to include both CINC score and any of these constituent parts in the same model. Thus,
for robustness, I run models with only CINC included in the model, and also models with each
individual component (as well as combinations thereof).

36

lllllllllllllllllllllllllllllllll19851990199520002005201020151000000150000020000002500000Total Word CountYearYearly total word count, with lowess smootherFigure 2.2: Average yearly word count, worldwide, across 1982-2014

Second, I include a measure of Nuclear capability, which in the post-World War 2 era serves as
a uniquely stark measure of a country’s coercive leverage, given the threat associated with Nuclear
weaponry. For this, I use Fuhrmann & Tkach’s (2015) data on Nuclear Latency. This data moves
beyond the binary classiﬁcation of nuclear powers and non-nuclear powers, and also includes the
ambition for nuclear power by coding countries with "at least a laboratory-scale enrichment or
reprocessing plant" and countries with "at least a pilot-scale enrichment or reprocessing plant".
Using this data, I create a 4-category variable composed of None, Either lab or pilot, Both lab and
pilot, and Full.

Measuring democracy is a trickier task. There is a robust literature on how to measure

37

183420480Average word count per year, 1982−2014Figure 2.3: Average yearly word count, NAM countries, across 1982-2014

Democracy, given the high dimensionality of the concept as well as the prioritization of diﬀerent
elements of democracy. Extensive work has been done to resolve these issues. Some work,
such as the highly-cited18 Polity score, the Freedom House scores (House, 2014) or Przeworski’s
(2000) scores 19, sets out to aggregate across various measures and indicators to create a single
index. Recently, building oﬀ from this has been a second, related method which uses statistical
models to both weight various attributes of democracy and to account for measurement error in
these estimates (Treier & Jackman, 2008); these methods have also been used across measures of
democracy themselves–as opposed to on the underlying attributes–to help "minimize the impact of

18But recently controversial; see Treier & Jackman (2008); Boese (2019)
19Also see Cheibub et al. (2010)

38

187320480Average word count per year, 1982−2014 (Non−Aligned Movement)Figure 2.4: Average yearly word count, EU countries, across 1982-2014

Figure 2.5: Average yearly word count, AU countries, across 1982-2014

39

1412020450Average word count per year, 1982−2014 (EU members only)413320480Average word count per year, 1982−2014 (African Union Countries)idiosyncratic errors that occur in individual measures, to take advantage of the level of agreement
between raters" across independent measures (Pemstein et al., 2010). Finally, a third method
embraces multi-dimensionality by explicitly breaking out democracy on multiple dimensions,
providing unique scores across these diﬀerent dimensions while also providing aggregated scores
in addition to breaking out dimensions (Lindberg et al., 2014), a method which has seen some
validation in the literature (Boese, 2019) as compared to others.

One advantage of the third method is that the explicit mechanism in question can be more
carefully measured. More speciﬁcally, the argument presented here is that democracies are most
optimally suited to incentivize leaders to cater politically to the diversity of preferences in their
respective populations. This is a limiting factor of more popular indices such as the Polity index,
which is primarily focused on the institutional structure behind the leader’s power (i.e. executive
recruitment, political competition, and executive independence).20 Using V-Dem data (Lindberg
et al., 2014), I use two measures to capture the institutional link between leader and constituent:
the electoral democracy index, which measures the extent of suﬀrage; freedom of civil society
to operate; cleanliness and fairness of elections; and most importantly whether "elections aﬀect
the composition of the chief executive of the country"–a crucial mechanism to linking constituent
to leader. The second indicator I use from V-Dem is the participatory democracy index, which
relatedly measures the extent to which there is maximal participation in the political process by the
population (including civil society engagement) and also participatory institutions such as direct-
democracy and sub-national governance. Finally, to be thorough, I also use the Uniﬁed Democracy
Scores (Pemstein et al., 2010) from the work on statistical measurement models, since this measure
provides a rigorous approach to compromising between using any single indicator (such as Polity or
Freedomhouse) and also having a single, uniﬁed measure to compare across countries. All results
are similar to those presented here.

20However, given its prevalence in the literature, I also use the Polity score as a robustness check.

40

2.2.1.2 Control variables

There are other characteristics, be they country-level or structural, that may be expected to contribute
to a country’s engagement with the UN General Assembly. For instance, while material capacity is
a variable of interest here, the ﬁve permanent members of the UN Security Council are also among
the most powerful countries in the world, and yet likely have diﬀerent incentives due to their vastly
superior position to leverage the UN Security Council. Whether this leads to less speech (due to
their capability to coerce favorable outcomes at the UNSC) or more speech (due to their heightened
interest in issues of the UN due to this position), I account for this as a separate characteristic. To
do this, I include binary indicators of the 3 Western powers of the Permanent-5 members of the
Security Council (The United States, Great Britain, and France), and a binary indicator for the 2
other Permanent-5 members (China and Russia). For a related reason, I also include an additional
binary measure of whether a country is a member of the UN Security Council at all–including the
rotating membership every 2 years. Finally, given that some countries simply may have unobserved
reasons to value international institutions over others–making them more prone to activity at IGOs
at all–I include a measure of a country’s general propensity to engage with international institutions
by using the count of Inter-Governmental Organizations to which a country is party Pevehouse
et al.’s (2004).

I also include variables to measure a country’s alignment with the UN’s organizational mission
in protecting peace and security, promoting human rights, and mitigating violence (Fortna, 2008;
Beardsley & Schmidt, 2012; Binder, 2015). For example, we might expect that countries involved a
conﬂict – particularly intense conﬂicts – would see diﬀerent levels of engagement with the UN. For
this, I use the Uppsala Conﬂict Data Program and Peace Reserach Institute of Oslo’s (UCDP/PRIO)
data on international conﬂict; these code all conﬂicts with greater than 25 battle deaths as conﬂicts,
with an indicator for whether a conﬂict crosses the 1000 battle death threshold. This allows me
to construct two variables, for conﬂict and for intense conﬂicts. I label any country which is an
actor (i.e. on side A, side B, or the location of the conﬂict) in a conﬂict within a given year as
experiencing a conﬂict. To capture alignment with the UN’s normative mission to promote human

41

rights, I include a measure of a country’s Human Rights record using Fariss’s (2014) latent variable
measure of Human Rights. Finally, I include a logit-transformed proportion of No or Abstain
votes that a given country made in a given year, out of all the resolutions in that year. This should
provide an additional measure of a country’s interest in the immediate mechanics of the UN General
Assembly in a given year.

As brieﬂy discussed above, one of the key variables of interest–the Composite Indicator of
National Capabilities–is composed of multiple, more ﬁne-grained measures of a country’s broad
material capacity. These are also important characteristics that are not only standard in the literature
to explaining political outcomes, but also have direct implications for more concretely understanding
material capacity. These include GDP Per Capita (Food & of the United Nations, 2019b)–which is
a measure of the more speciﬁc economic material capacity–as well as Population (Food & of the
United Nations, 2019a), which is a measure of both human capacity as well as a proxy for the
diversity of preferences within the nation, which overlaps with democracy as a key variable as well.
In the framework of this paper, both should be expected to have a very strong relationship with a
country’s engagement with the UN General Assembly. I do not include Population and CINC in
the same models, since the latter is a composite including the former and thus collinearity is severe.
As Table 2.1 shows, the correlation between CINC and population is extreme, at around 0.94–I
include Population in all models that do not contain CINC, with the understanding that it is also a
key measures of material capacity.

Finally, despite the relative independence of UNGA speeches, there is a unique and important
source of interdependence between countries in terms of their speeches at the UNGA. It is relatively
common practice at the UNGA for blocs of countries – such as the Arab League, European Union,
Non-Aligned Movement, etc. – to deliver foreign policy speeches as a group, which by deﬁnition
results in word counts being extremely similar for those countries. These groups are often formed
on the very fact that these countries share common foreign policy interests, and so they must
be accounted for when modeling individual countries’ decisions to speak. Therefore, during the
process of parsing and organizing speeches, I identiﬁed speeches which were given "on behalf of"

42

an identiﬁable group, or those which are given by a country "in their capacity as" the chairperson
or leader of one of these groups. In this analysis, I include a binary variable for whether a country
is a member of the EU, Non-Aligned Movement, African Union, or ASEAN. I also include a mean-
centered variable capturing the word count delivered as part of a group in the models of overall
word count. Finally, to further account for interdependence, I include a cubic polynomial of Bailey
et al.’s (2015) ideal point measure of UNGA voting aﬃnity, to further account for unobserved
hetereogeneity and interdependence.21

2.2.1.3 The dependent variable

The outcome of interest is the number of words spoken at the UN General Assembly by an individual
country in a given year. The variable is normally distributed with a slight right skew; Figure 2.6
provides a distribution of individual countries’ speeches across the full 1982-2014 time period.
The distribution is slightly skewed right, but not substantially; yearly distributions are similar to the
average. There are two possible approaches to modeling this, each with compelling justiﬁcation.
First, while words are superﬁcially an ordinal count, in that there cannot be a decimal number of
words, they are not independent of one another; in fact, they are explicitly dependent–and more
precisely, they are co-ocurring. Words are a mechanism to achieve communication of an underlying
concept: two words, be they in a single sentence or a single paragraph, co-occur by design during
the process of translating an unspoken latent concept into spoken communication.22 Treating these
as counts in a statistical sense would be nonsensical, since the volume of words in totality can be
treated as a single, uniﬁed realization drawn from a random distribution across more eﬃcient (i.e.
communication in fewer words) or less eﬃcient (i.e. communication in more words) possibilities
in communication. Thus, we should instead treat word counts as being centered on some average
21The results are substantively similar when I instead break these measures into 5 equal categories,

and include these as categorical variables.

22Indeed, the acute reader would notice this is precisely the logic of topic models and (to some
extent) word embeddings in natural language processing, whereby words are observed instances of
some unobserved, latent concepts that can either be derived as a latent variable (Blei et al., 2003)
or measured in parallel to other concepts in word embeddings (Collobert et al., 2011).

43

Figure 2.6: Average yearly word count per individual country, 1982-2014

level of eﬃciency in an approximately normal distribution, such that some speakers will randomly
be more or less eﬃcient (all else being equal) in their communication. Here, we would expect that
given this random distribution of eﬃciency, changes in the mean of the distribution would be driven
by more content to communicate–and thus capture the political mechanism under analysis here.

On the other hand, as others have pointed out in the literature, such correlated counts still need
not mean we can model them using a normal distribution. For example, writing on the topic of
political violence, Krain (1998) points out that models like OLS (and the Normal distribution more
broadly) would allow for nonsensical negative values–which is relevant for speech, since word
counts cannot take on negative values. Further, it assumes that a single unit change is the same
across units, whether it be the diﬀerence between 100 and 200 words, or the diﬀerence between
10,000 and 10,100 words. This is also problematic, since in theory, as word counts increase, the
importance of adding more words would decrease. Finally, they argue, while it is true that count
models in an ideal form assume independence, "correlated events create situations of overdispersion
or underdispersion, which can be accounted for by more complex count models, such as the negative
binomial or zero modiﬁed count models" (Krain, 1998, p.145, citing Liao (1994); Long (1997)).

44

−2−10123450.00.10.20.30.4Mean−centered Word CountN = 6130   Bandwidth = 0.1562DensityThus, while there are compelling theoretical arguments for the data-generating process of text to
be treated as a normally distributed variable centered on some average speech amount, there are
also statistical reasons to believe that event-count models such as the Negative Binomial would be
suﬃcient, and in some ways may be superior especially due to both non-negative values and the
changing importance of one-unit changes.

Granting these diﬃculties, I model word counts using a negative binomial distribution. However,
all models are substantively similar in alternative models.23 Brieﬂy, the negative binomial model
is a generalization of the poisson model, which models the probability of count  ocurring:

( = ) =

   (−)


!

(2.1)

where the parameter  is modeled with a natural log transformation  () = 

 , where
 is a vector of coeﬃcients theorized to explain the variance in the expected count–in this case,
the number of words spoken. While the poisson model assumes that both the expected value and
the variance are both equal to , this assumption is often violated (in this paper, for example,
the variance in the data is over 3000 times the mean); thus, the negative binomial model adds a
gamma-distributed variance parameter  to account for overdispersion (or underdispersion).

While no transformation is required for the dependent variable, since the negative binomial
accounts for the changing importance of unit changes, some of the independent variables in the
model are highly skewed. Namely, CINC score; GDP Per Capita; and Total Population are all
extremely right-skewed, since material capacity is highly unequal in its distribution across the
international system. Thus, I log-transform these three variables in the estimated models. Finally,
I center the count of IGO memberships at its mean to facilitate interpretation. All analysis and
23Namely, I subject these results to additional robustness checks, including OLS, Robust Re-
gression (Western, 1995)–which uses iterative least squares to downweight outliers–as well as
bootstrapped robust regressions (Zeileis, 2004, 2006) and OLS with AR1-panel-corrected standard
errors (Kashin, 2014). Additionally, I also estimate a model with a lagged dependent variable to
account for any temporal stickiness and interdependence between yearly engagement. The results
are substantively similar.

45

model estimation was done using R.24

2.3 Empirical patterns

Bivariate relationships and correlations between key variables provide strong support for the
main claims in this paper. Table 2.1 provides a table of bivariate correlations between the Uniﬁed
Democracy scores, V-Dem components, natural-log transformed CINC scores, non-transformed
CINC scores, Word Counts at the UN General Assembly, Population, UN Security Council mem-
bership, and major power status (P2 and P3). The correlation between the log-transformed CINC
and Word count is around 0.43, a remarkably high correlation; similarly, the correlation between
Democracy and word count is about 0.25. Figure 2.7 provides a loess-smoothed graphical relation-
ship of CINC score against word count, across the entire time period; Figure 2.8 provides a similar
plot across Uniﬁed Democracy scores. Descriptively, there is a clearly discernible propensity for
powerful and democratic countries to speak more at the UN General Assembly–across all measures.

24More speciﬁcally, the MASS package (Venables & Ripley, 2002) to estimate negative binomial
models, the "visreg" (Breheny & Burchett, 2015) package in some cases to visualize the results,
and the "texreg" (Leifeld, 2013) package to output tables. Robust regression was done using the
"rlm" package (Zeileis, 2004, 2006).

46

centWTwc
cinc
logcinc
UDSmedian
v2x_partipdem
v2x_polyarchy
v2x_libdem
v2x_delibdem
v2x_egaldem
logtpopFAO
loggdppcFAO
UNSCmem
P3
P2

centWTwc
1.00
0.26
0.40
0.30
0.34
0.30
0.34
0.34
0.36
0.32
0.38
0.20
0.20
0.09

cinc
0.26
1.00
0.53
0.05
0.07
0.05
0.07
0.10
0.07
0.51
0.13
0.46
0.40
0.65

0.40
0.53
1.00
0.05
0.11
0.06
0.08
0.11
0.08
0.91
0.23
0.32
0.26
0.27

logcinc UDSmedian v2x_partipdem v2x_polyarchy v2x_libdem v2x_delibdem v2x_egaldem logtpopFAO loggdppcFAO UNSCmem
0.20
0.46
0.32
0.10
0.13
0.11
0.13
0.14
0.13
0.30
0.15
1.00
0.43
0.35

0.30
0.05
0.05
1.00
0.92
0.94
0.93
0.92
0.91
-0.03
0.59
0.10
0.18
-0.09
Table 2.1: Bivariate correlations with UNGA Word Count

0.36
0.07
0.08
0.91
0.95
0.94
0.97
0.96
1.00
-0.05
0.67
0.13
0.21
-0.08

0.38
0.13
0.23
0.59
0.59
0.55
0.63
0.61
0.67
-0.08
1.00
0.15
0.21
-0.01

0.32
0.51
0.91
-0.03
0.03
-0.00
-0.02
0.02
-0.05
1.00
-0.08
0.30
0.21
0.27

0.34
0.07
0.11
0.92
1.00
0.97
0.97
0.96
0.95
0.03
0.59
0.13
0.22
-0.11

0.30
0.05
0.06
0.94
0.97
1.00
0.97
0.97
0.94
-0.00
0.55
0.11
0.20
-0.11

0.34
0.07
0.08
0.93
0.97
0.97
1.00
0.98
0.97
-0.02
0.63
0.13
0.22
-0.11

0.34
0.10
0.11
0.92
0.96
0.97
0.98
1.00
0.96
0.02
0.61
0.14
0.21
-0.08

P3
0.20
0.40
0.26
0.18
0.22
0.20
0.22
0.21
0.21
0.21
0.21
0.43
1.00
-0.02

P2
0.09
0.65
0.27
-0.09
-0.11
-0.11
-0.11
-0.08
-0.08
0.27
-0.01
0.35
-0.02
1.00

47

Figure 2.7: UNGA word count vs CINC

Table 2.2 tabulates coeﬃcients from estimated negative binomial models, across diﬀerent model
speciﬁcations. The positive relationship between power and democracy–under both participatory
and polyarchy deﬁnitions (Lindberg et al., 2014)–remains robust to the inclusion of a battery of
control variables. Table 2.3 provides the Incident Response Rates (i.e. exponentiated coeﬃcients)
with 95% conﬁdence intervals for key variables in this paper.25 A change from no Nuclear capability
to having either a lab or a pilot program is related to between a 13% - 33% increase (95% CI) in
the expected amount of speech, and a change from no capability to full nuclear capability is around
31% to 61% increase (95% CI). Figure 2.9 plots the predicted values of speech across levels of
Nuclear capability from model 1; Figure 2.10 and Figure 2.11 provide these plots for CINC score
and V-Dem Polyarchy measures, including a plot with both log-transformed and un-transformed
CINC scores. Figure 2.12 plots predictions for V-Dem Participatory measures from model 2.

25Brieﬂy, given that the regression coeﬃcients are the log of expected counts, and the diﬀerence
between unit  and  + 1 can be given as  (+1) −  (), yielding a ratio 
. The
IRR, then, exponentiates this ratio, providing interpretability as log-odds.


(cid:16) +1

(cid:17)

48

024−12−8−4log CINC ScoreMean centered word count1982 to 2014Bivariate Loess−smoother between power and UNGA word countFigure 2.8: UNGA word count vs Democracy

Figure 2.9: Predicted word count (with 95% CI), UNGA word count (cent) vs Nuclear capability

49

024−2−1012Universal Democracy ScoreMean centered word count1982 to 2014Bivariate Loess−smoother between Democracy and UNGA word count600070008000900010000110000None1LabOrPilot2LabAndPilot3FullNucLatWTwcModel 1
9.488∗∗∗
(0.083)
0.100∗∗∗
(0.006)
0.243∗∗∗
(0.040)

0.387∗∗∗
(0.009)
−0.017
(0.025)
−0.430∗∗∗
(0.072)
−0.000
(0.080)
−0.300∗∗∗
(0.032)
0.078∗∗∗
(0.023)
0.244∗∗∗
(0.043)
0.077∗
(0.036)
0.033∗
(0.013)
0.060∗∗∗
(0.008)
−0.041
(0.034)
0.052
(0.035)
0.203∗∗∗
(0.042)
0.093∗∗
(0.036)
0.372∗∗∗
(0.052)
0.016∗
(0.008)
0.002
(0.002)
3884
Y
Y
Y

(Intercept)

log(CINC)

vDEM polyarchy

vDEM partipdem

Group WTwc

UNSC Seat

P3

P2

EU Member

NAM Member

AU Member

ASEAN Member

cent IGO Count

log(GDPpc)

UCDP conﬂict

UCDP intense conf

NucLat 1: LabOrPilot

NucLat 2: LabAndPilot

NucLat 3: Full

HumRights

logit(NoOrAbstVoteProp)

Num. obs.
Year Dummy Vars
Regions
UNGA IP Polynomial
AIC
Log Likelihood
Deviance
∗∗∗  < 0.001, ∗∗  < 0.01, ∗  < 0.05

Model 2
9.496∗∗∗
(0.083)
0.099∗∗∗
(0.006)

0.384∗∗∗
(0.055)
0.384∗∗∗
(0.009)
−0.018
(0.025)
−0.427∗∗∗
(0.072)
0.022
(0.080)
−0.299∗∗∗
(0.032)
0.085∗∗∗
(0.023)
0.243∗∗∗
(0.043)
0.084∗
(0.036)
0.029∗
(0.013)
0.057∗∗∗
(0.008)
−0.039
(0.034)
0.047
(0.035)
0.203∗∗∗
(0.042)
0.097∗∗
(0.036)
0.369∗∗∗
(0.052)
0.014
(0.008)
0.002
(0.002)
3884
Y
Y
Y

73936.960
73925.811
-36915.480 -36909.906
3976.370
3976.150

Table 2.2: Negative Binomial Regression, UNGA word counts, yearly

50

(Intercept)
log(cinc)+
vDEM_polyarchy+
vDEM_partipdem++
loggdppc+
NucLat1LabOrPilot+
NucLat2LabAndPilot+
NucLat3Full+
+Based on Model 1; ++Based on Model 2

exp(Est.)
13196.79
1.11
1.27
1.47
1.06
1.22
1.10
1.45

11208.95 15537.16
1.12
1.38
1.64
1.08
1.33
1.18
1.61

97.5% z val.
113.90
16.80
6.08
6.99
7.64
4.78
2.60
7.10

2.5%

1.09
1.18
1.32
1.05
1.13
1.02
1.31

p
0.00
0.00
0.00
0.00
0.00
0.00
0.01
0.00

Table 2.3: Incident Response Ratios for variables of interest

Figure 2.10: Predicted word count (with 95% CI), UNGA word count (cent) vs CINC

51

5000750010000125000.000.050.100.150.20cincWTwcFigure 2.11: Predicted word count (with 95% CI), UNGA word count (cent) vs VDem Polyarchy

Figure 2.12: Predicted word count (with 95% CI), UNGA word count (cent) vs VDem
Participatory

52

6000700080000.000.250.500.75v2x_polyarchyWTwc60007000800090000.00.20.40.60.8v2x_partipdemWTwc2.3.1 Evaluating mechanisms

There are two mechanisms under discussion here, each captured respectively under the empirical
characteristics of material capacity and political incentives. Both have additional considerations
and potentially clarifying elements for the mechanism translating them to UN General Speeches.
This section provides some insight into these more speciﬁc elements.

2.3.1.1 Material capacity

First, I argue that countries with higher levels of material capacity will have the capability to
become involved with a larger number of international issues, and thus will have a more diverse set
of issues and political perspectives to address within their speeches. There are at least two possible
mechanisms within this argument. The ﬁrst mechanism is that the most powerful countries have
extensive coercive power, both at the UN Security Council (Voeten, 2001) to leverage outside
options to coerce others to bend toward their will, and at the UN General Assembly (Carter &
Stone, 2015) to explicitly (or implicitly via threat) buy votes. Here, the goal of speech is to
reach politically relevant audiences who are interested in a given issue–and material capacity
incentivizes these countries to voice their opinions on a larger set of issues. This mechanism
is inherent to material capacity, since these countries have both the military and the economic
leverage to pressure countries on a variety of issues. This mechanism is also diﬃcult to test, since
the observable implications of coersion are scarce: the observed behavior of countries would be
the result of this leverage, and thus using observable data to test it is extremely diﬃcult (Gartzke,
1999). In a sense, this can be treated as a quasi-null hypothesis.

The second mechanism is that these powerful countries have a propensity to involve themselves
in global aﬀairs, and as a result these underlying propensities drive observable involvement, which
then leads to increased speech. This is much more straightforward to test. There is some preliminary
support for this hypothesis: as Table 2.2 shows, the number of Inter-Governmental Organizations
to which a country is party is realted to an increased amount of speech at the UN General Assembly
(0.5% to 6% per unit increase (95% CI)). If the ﬁnding about material capacity is also explained

53

by the propensity to engage, the positive relationship between material capacity and UN General
Assembly speech should be more pronounced for those countries with an already-demonstrated
propensity to engage. To put it diﬀerently, since weaker countries have a much lower ceiling in their
capacity to engage–whether their potential capacity via coercive power, or expressed capacity via
actual involvement abroad–the diﬀerence between highly engaged and less engaged weak countries
would be smaller than the diﬀerence between highly engaged and less engaged strong countries.

Empirically, this suggests an interactive relationship between material capacity and the number
of IGOs to which a country is party. Note that this also corresponds well to Berry et al.’s (2012)
guidance on forming theoretically driven interaction eﬀects, where they point out that "...logically,
all interactions are symmetric...In other words, if Z modiﬁes the eﬀect of X on Y, then X must
modify the eﬀect of Z on Y" (Berry et al., 2012, p.653, citing (Brambor et al., 2006; Kam &
Franzese, 2007)).
the
relationship between material capacity and UNGA speech should be ampliﬁed when countries are
more embedded in the system (given the increase in propensity to engage), and that the relationship
of embeddedness and UNGA speech should increase as material capacity increases, due to the
higher ceiling of possible issues a country can aﬀect at all. To test this, I use a multiplicative term
between these two variables.

In this case, the two theorized, complementary relationships are clear:

Table 2.4 provides output of the negative binomial regression on UNGA word counts, including
an interaction between individual measures of material power and the number of IGOs to which
a country is party. The results are mixed. The results on the interaction between CINC and IGO
count are generally positive, but the magnitude changes drastically depending on whether GDP per
capita is included in the model. Figure 2.13 plots the change in marginal eﬀect for both IGO and
CINC from Model 2. This is also complicated by, as discussed above, the fact that GDP per capita
is actually a constituent part of the CINC score. While the correlation between the two, as given in
Table 2.1, is not as unusually high as it is for population, the results in Model 2 show that removing
GDP Per Capita, and isolating only CINC score, the interaction term is in the expected direction.
However, GDP per capita is the opposite direction: Model 3 shows that the interaction between

54

Figure 2.13: Marginal eﬀect of CINC score across IGO membership, and vice versa

GDP Per capita and IGO membership is negative, which contradicts the propensity to engage
mechanism (also see Figure 2.14. This may lend some credibility to the more speciﬁc coercion
argument, since Nuclear Latency (model 4) shows similar patterns as CINC score. In other words,
there is some evidence that the propensity to actually engage with international issues ampliﬁes the
relationship between coercive power and the level of speech at the UN General Assembly, but there
is little to falsifying evidence that it ampliﬁes the relationship between economic power and speech
at the UN General Assembly. Thus, the speciﬁc mechanism at play seems to fall back to more
traditional explanations of international politics:
if international politics ultimately reﬂects the
will of the militarily powerful, speeches at the UN General Assembly reﬂect some combination of
speciﬁcally this military power distribution (i.e. opportunity) and the actual willingness to engage
with international political issues in the ﬁrst place.

2.3.1.2 Political incentives

Second, I argue that countries with a stronger connection between leader and constituent will
incentivize leaders to speak more to cover all the angles of a given issue, as well as to address

55

lllllllllllllllllllllllll0.040.080.120.16−2−10123Mean−centered IGO membershipMarginal Effect of log(CINC)lllllllllllllllllllllllll−0.050.000.050.100.150.20−12.5−10.0−7.5−5.0−2.5log(CINC)Marginal Effect of Mean−centered IGO membershipModel 3
8.754∗∗∗
(0.072)

Model 4
8.765∗∗∗
(0.077)

0.331∗∗∗
(0.041)
0.226∗∗∗
(0.039)
0.372∗∗∗
(0.009)
0.050∗
(0.024)
−0.194∗∗
(0.063)
0.444∗∗∗
(0.062)
−0.321∗∗∗
(0.033)
−0.001
(0.022)
0.121∗∗
(0.039)
0.075∗
(0.034)
0.068∗∗∗
(0.008)
0.053∗∗∗
(0.012)
−0.037
(0.022)
−0.051∗∗∗
(0.007)
0.005∗
(0.002)

−0.019∗∗∗
(0.005)

0.154∗∗∗
(0.011)
0.186∗∗∗
(0.041)
0.370∗∗∗
(0.010)
0.012
(0.026)
−0.071
(0.099)
0.245∗∗
(0.085)
−0.356∗∗∗
(0.033)
0.009
(0.023)
0.079
(0.043)
0.189∗∗∗
(0.037)
0.065∗∗∗
(0.008)
0.013
(0.014)
0.007
(0.024)
−0.045∗∗∗
(0.008)
0.006∗∗
(0.002)

0.757∗
(0.310)
0.494
(0.256)
1.414∗∗∗
(0.214)
−0.058
(0.037)
−0.027
(0.027)
−0.145∗∗∗
(0.028)
3884
Y
Y
Y

Model 1
9.459∗∗∗
(0.084)
0.106∗∗∗
(0.006)
0.064∗
(0.030)
0.291∗∗∗
(0.040)
0.386∗∗∗
(0.009)
−0.014
(0.025)
−0.231∗∗∗
(0.067)
0.313∗∗∗
(0.065)
−0.312∗∗∗
(0.033)
0.090∗∗∗
(0.023)
0.267∗∗∗
(0.041)
0.013
(0.036)
0.062∗∗∗
(0.008)
0.027∗
(0.013)
−0.010
(0.023)
0.009
(0.008)
0.002
(0.002)
0.004
(0.004)

Model 2
9.925∗∗∗
(0.060)
0.103∗∗∗
(0.006)
0.152∗∗∗
(0.029)
0.301∗∗∗
(0.041)
0.388∗∗∗
(0.009)
0.002
(0.026)
−0.311∗∗∗
(0.070)
0.356∗∗∗
(0.065)
−0.333∗∗∗
(0.034)
0.077∗∗∗
(0.023)
0.171∗∗∗
(0.040)
0.080∗
(0.036)

0.014
(0.013)
0.022
(0.023)
0.039∗∗∗
(0.007)
0.002
(0.002)
0.011∗∗
(0.004)

(Intercept)

log(CINC)

Mean-Cent IGO

vDEM polyarchy

Mean-Cent Group WC

UNSC mem

P3

P2

EU

NAM

AU

ASEAN

log(GDPpc)

Num Conf

Intense Conf

Hum Rights

logit(Prop NoVotes)

log(CINC):centIGO

log(GDPpc):centIGO

NucLat1 - LabOrPilot

NucLat2 - LabAndPilot

NucLat3 - Full

log(GDPpc):NucLat1

log(GDPpc):NucLat2

log(GDPpc):NucLat3

3884
Y
Y
Y

Num. obs.
Year Dummy Vars
Regions
UNGA IP Polynomial
AIC
Log Likelihood
Deviance
∗∗∗  < 0.001, ∗∗  < 0.01, ∗  < 0.05

4070
Y
Y
Y

4214
Y
Y
Y

77768.560

74001.025
74194.798
-36949.512 -38834.280 -40245.959 -37042.399
3981.565
3977.679

80595.918

4176.091

4317.480

Table 2.4: UNGA Word Count (mean-centered), with Power*IGO interactions

56

Figure 2.14: Marginal eﬀect of Log GDP Per Capita across IGO membership, and vice versa

alternative issues that may otherwise not recieve coverage. One critique is that the Democracy
ﬁnding may simply be an artifact of a propensity towards Liberal values: simply put, it may be
argued that democracy does not only capture the domestic audiences to whom a leader wishes
to speak, but rather that it captures the underlying Liberal values that predispose a country to
engaging in a Liberal institution like the UN. To clarify this, we can also turn to IGO membership
as an opportunity: IGO membership provides, among other things, an internationalized measure of
Liberal values; thus, to validate this critique we would expect that the highest propensity to speak
should be democratic countries who are members of large amounts of IGOs, since this would
maximize a given country on both domestic and international expressions of Liberal values.

Table 2.5 provides output of the negative binomial regression on UNGA word counts, including
an interaction between two diﬀerent V-Dem measures of democracy, and the number of IGOs to
which a country is party. The results cast doubt on the Liberal Values critique: as shown in the
perspective plot in Figure 2.15, the highest propensity to speak actually comes from democratic
countries who are party to the least number of IGOs. This isolates the relationship as being one
more driven by Democracy speciﬁcally, rather than the broader underlying Liberal values that might

57

lllllllllllllllllllllllll0.000.050.100.15−2−10123Mean−centered IGO membershipMarginal Effect of log(GDPpc)lllllllllllllllllllllllll0.10.20.34681012log(GDPpc)Marginal Effect of Mean−centered IGO membershipFigure 2.15: 2D perspective plot, UNGA word count vs Democracy and IGO membership

predispose a country to IGO engagement. Figure 2.16 provides marginal eﬀects for Democracy
and IGO membership, showing that contrary to the expectations of a Liberal Values hypothesis,
among democracies the relationship of IGO membership to UNGA word count is actually the
most negative. This also suggests that the UN may be one of many forums in the world for
leaders to speak to key constituents, with those leaders who get the most marginal beneﬁt from this
communication–those with the least number of alternative forums (in this case, alternative IGOs)
to leverage–employing it the most. This particular question of choosing between alternative forums
for communication may be an interesting avenue for future research.

2.4 Linking speech to interest on relevant UN issues

If UN General Assembly speeches reﬂect the political incentive to inﬂuence issues relevant to
the UN, we should expect that countries who speak more at the UN General Assembly will also
reﬂect an interest in the issues that the UN as an institution addresses. While establishing political
and material bases to speech is valuable, this is ultimately incomplete without a link to a tangible,
otherwise costly behavior that directly implicates a political interest. To that end, I measure whether

58

v2x_partipdem0.20.40.60.8centIGOmem−2−10123f(v2x_partipdem, centIGOmem)8.68.89.09.2Model 1
9.513∗∗∗
(0.084)
0.111∗∗∗
(0.006)
0.253∗∗∗
(0.040)
0.144∗∗∗
(0.022)
0.384∗∗∗
(0.009)
−0.008
(0.025)
−0.141∗
(0.065)
0.231∗∗∗
(0.066)
−0.252∗∗∗
(0.034)
0.082∗∗∗
(0.023)
0.215∗∗∗
(0.041)
−0.018
(0.035)
0.068∗∗∗
(0.008)
0.033∗
(0.013)
−0.011
(0.023)
0.015
(0.008)
0.001
(0.002)
−0.198∗∗∗
(0.032)

(Intercept)

log(CINC)

vDEM polyarchy

Cent IGO

Cent Group WC

UNSC mem

P3

P2

EU

NAM

AU

ASEAN

log(GDPpc)

Num Conf

Intense Conf

Hum Rights

logit(Prop NoVotes)

vDEMpoly:centIGO

vDEM partipdem

vDEMpartip:centIGO

Model 2
9.501∗∗∗
(0.083)
0.110∗∗∗
(0.006)

0.131∗∗∗
(0.019)
0.379∗∗∗
(0.009)
−0.010
(0.025)
−0.102
(0.065)
0.242∗∗∗
(0.065)
−0.240∗∗∗
(0.034)
0.088∗∗∗
(0.023)
0.214∗∗∗
(0.041)
−0.013
(0.035)
0.067∗∗∗
(0.008)
0.030∗
(0.013)
−0.011
(0.023)
0.012
(0.008)
0.001
(0.002)

0.444∗∗∗
(0.055)
−0.291∗∗∗
(0.040)
3884
Y
Y
Y

3884
Y
Y
Y

Num. obs.
Year Dummy Vars
Regions
UNGA IP Polynomial
AIC
Log Likelihood
Deviance
∗∗∗  < 0.001, ∗∗  < 0.01, ∗  < 0.05

73964.551
73937.511
-36931.275 -36917.755
3976.441
3976.973

Table 2.5: UNGA Word Count, with V-Dem*IGO interactions

59

Figure 2.16: Marginal eﬀects (with 95% CI), vDEM Particip and IGO membership

higher levels of engagement with the UN General Assembly provides explanatory power towards
whether a country actually has some investment in the UN as an institution overall. To do this,
I turn to seminal work by Kathman & Melin (2017), who model why particular countries devote
troops to UN Peacekeeping Operations.

In their paper, Kathman & Melin (2017) argue that countries deploy troops to UN Peacekeeping
Operations, in part, to areas that "focus the foreign and security policy attention of states". They
measure this via the presence of an interstate rivalry, showing that countries are more likely to
contribute troops where a rivalry is present. If UN General Assembly speeches reﬂect the presence
of foreign and security policy attention of states on the issues relevant to the UN, we should expect
that countries who deliver more speech at the UN General Assembly will be more likely, in general,
to contribute troops–and also contribute more of them (all else equal).
I replicate Kathman &
Melin’s (2017) work to determine whether the addition of UN engagement shows an additional,
independent relationship with this decision. Kathman & Melin (2017) use a Zero-Inﬂated negative
binomial model, which simultaneously models the decision to send UN peacekeepers at all, as well
as the count of troops deployed given that decision. In this case, we should expect that engagement

60

lllllllllllllllllllllllll−0.50.00.51.01.5−2−10123Mean−centered IGO membershipMarginal Effect of VDEM Participwith the UN should be related to both. To do this, Kathman & Melin (2017) draw from original
data in Kathman (2013) on troop contributions to all UN PKOs between 1992-2010, which falls
well within the time period under study in this paper.

Descriptively, there is a fairly robust correlation between UN General Assembly engagement
and the number of troops deployed to UN Peacekeeping operations. Figure 2.17 provides a Loess-
smoothed plot over the bivariate relationship between UNGA engagement and the number of troops
sent, showing a strong positive relationship.

Table 2.6 provides the output from the Zero-Inﬂated negative binomial model from Kathman
& Melin (2017), replicating both of their original analyses and adding in UNGA word counts. To
quote from Kathman & Melin (2017, 158), "...the interpretation of coeﬃcient directions in the
[ﬁrst stage of the model] is diﬀerent from a standard regression model interpretation, as positive
coeﬃcients indicate an increased likelihood of contributing ’nothing’ (i.e., a decreased likelihood
of contributing ’something’), whereas a negative coeﬃcient indicates a decreased likelihood of
contributing ’nothing’ (i.e., an increased likelihood of contributing ’something’)." In other words,
the ﬁrst stage models the probability of contributing no troops, while the second stage models the
count of troop contributions among countries who do end up contributing troops.

The results show support for the claim that the amount of UN General Assembly speech is
measuring a level of interest in the issues relevant to the UN more broadly. Replicating Kathman
& Melin’s (2017) methodology for plotting predictions from the model, Figure 2.18 plots the pre-
dicted counts of Troop contributions to UNPKO operations across the centered and scaled value of
word count at the UN General Assembly. Movement from speaking one standard deviation below
the yearly mean to speaking one standard deviation over the mean yields approximately a 400%
increase in the predicted count of troops devoted to UNPKO operations, holding other variables
constant at their means.26

26These results also hold to almost identical results when we exclude Pakistan, India, Bangladesh,
and Egypt, which are the 4 highest troop contributors. As shown in Figure 2.2, Pakistan, India, and
Egypt in particular have noticeably high levels of speech overall, hence the value of validating the
robustness of this ﬁnding to their removal from the dataset.

61

Figure 2.17: Loess smooth bivariate relationship, UNGA engagement and troop contribution

Figure 2.18: Predicted Counts of UNPK troops across word count

2.5 Discussion and conclusion

In summary, powerful countries with strong political accountability structures tend to speak the
most at the UN General Assembly. The impact of material capacity is most present among countries
that are highly integrated into the international system; this suggests that the increased propensity

62

(1)

troop count

(2)

troop count

(3)

troop count

(4)

troop count

Second Stage

Rivalry

Yrs Since last Coup

Yrs Since last Coup2

Lag Total Troops

UNGA WordCount

Maj Civil War

Total UN Missions

Democracy

Autocracy

Constant

First stage

0.238
(0.121)
0.0340
(0.0145)
-0.000521
(0.000228)
0.000790
(0.000140)

0.218
(0.125)
0.0303
(0.0144)
-0.000499
(0.000231)
0.000797
(0.000148)

0.136
(0.0606)

5.134
(0.251)

5.144
(0.248)

Lag Total Troops

UNGA WordCount

-6.476
(1.300)

Lagged contribute troops

-6.008
(0.749)
-0.887
(0.124)

Year

Year2

Year3

Democracy

Autocracy

Total UN Missions

Maj Civil War

GDP Per Capita

Americas

Europe

Africa

Middle East/North Af

Constant

lnalpha

AIC
log-likelihood
Standard errors in parentheses

2.409
(0.113)
0.916
(0.126)
16828.1
-8406.0

2.419
(0.114)
0.899
(0.104)
16483.5
-8231.8

0.226
(0.115)
0.0297
(0.0137)
-0.000421
(0.000214)
0.000859
(0.000168)

0.201
(0.124)
0.0275
(0.0135)
-0.000416
(0.000214)
0.000858
(0.000174)

0.110
(0.0569)
0.201
(0.301)
-0.104
(0.0190)
-0.108
(0.116)
0.101
(0.207)
6.946
(0.334)

-5.379
(0.344)
-0.921
(0.111)
0.476
(0.400)
0.816
(0.127)
-0.0903
(0.0140)
0.00263
(0.000446)

0.245
(0.245)
1.555
(0.421)
0.158
(0.0392)
0.522
(0.442)
-0.0646
(0.121)
0.780
(0.347)
0.500
(0.373)
0.206
(0.327)
0.664
(0.489)
-1.714
(1.086)
0.801
(0.0964)
16073.1
-8009.6

0.135
(0.280)
-0.109
(0.0186)
-0.0493
(0.110)
0.160
(0.205)
6.980
(0.342)

-5.592
(0.343)

0.426
(0.365)
0.776
(0.126)
-0.0857
(0.0141)
0.00252
(0.000446)

0.143
(0.235)
1.497
(0.395)
0.159
(0.0374)
0.405
(0.428)
-0.111
(0.113)
0.513
(0.342)
-0.270
(0.348)
0.111
(0.334)
0.0676
(0.462)
-0.917
(1.046)
0.796
(0.102)
16398.3
-8174.1

Table 2.6: Replication of Kathman and Melin (2017)

63

to speak is not due to integration in the system alone, but rather about the way material capacity
is linked to an underlying interest in being involved with international politics more broadly. This
has implications for the way we understand UN General Assembly speeches: namely, that what
we observe is disproportionately composed of countries with an already existing interest in world
issues, and who have some capacity to inﬂuence them. It also contradicts the conventional wisdom
that these types of speeches provide no beneﬁt to countries: under this scenario, we would expect
the impact of material capacity to be negative among those countries with the most interest in
inﬂuencing international politics, since those with high material capacity would have suﬃcient
private avenues to pursue their political interests. These ﬁndings show strong evidence that when
leaders speak at the UN General Assembly, they do so with an eye towards politically relevant
audiences to gain some speciﬁc beneﬁts toward the goal of inﬂuencing politics. Future research
should dig deeper into understanding the mechanisms of how the content of these speeches translate
into political mobilization toward inﬂuence–for example, ﬁltration and/or ampliﬁcation through the
media (Baum & Groeling, 2009; Baum & Potter, 2008) or the role of justiﬁcations on political
accountability (Levendusky & Horowitz, 2012; White & Ypi, 2011; Habermas, 1984).

The role of political institutions connecting relevant audiences to political leaders is also further
evidence that leaders can gain some political beneﬁts from these types of speeches, and contradicts
the null hypothesis that these speeches ultimately do not impact (or are not expected to impact) a
leader’s political success. The impact of political institutions on the incidence of speeches seems
directly linked to the opportunity to actually speak at all, since leaders who are less involved in
other institutions (where they can ﬁnd a forum to communicate) will be most impacted by these
political incentives. Descriptively, as shown in autorefﬁg:wcovertime, there is some evidence that
the amount of speech in general is growing over time. Given questions on whether democratization
is decreasing in the world (Lührmann & Lindberg, 2019; Mechkova et al., 2017; Diamond, 1996),
there may be opportunities to more concretely link international audiences to political leaders. This
could include mechanisms of international trade or other such links, be it vulnerability to external
leverage (Levitsky & Way, 2005) or the political decision to exchange sovereignty for economic and

64

military security (Lake, 2009). There may be additional patterns to be understood on the impact
of international or domestic political incentives on leaders propensity to speak publicly on issues,
which may be an area for future work.

Overall, these ﬁndings provide insight into a side of international institutions that, especially
in the empirical literature, has been relatively understudied: patterns of engagement by individual
member countries. It also provides a bridge between the coordination function of an institution
and the diplomatic, communicative function of the public forum contained therein. The question,
for example, of whether the provision of a highly visible public forum helps or hinders eventual
foreign policy cooperation through the institution is worth pursuing, given the politics surrounding
public speech presented here. This link between communication and institutions has far reaching
potential, given the proliferation of institutions in the international system over the past few decades,
as well as the increasing visibility–via mass media and the internet–of public speeches to potentially
inﬂuential elites at home and abroad.

65

CHAPTER 3

COMPOSITIONAL SPECIFICITY AND DIFFUSENESS OF SPEECHES ACROSS

WORLD ISSUES

Every year, nations of the world gather at the United Nations General Assembly (UNGA) to discuss
pertinent issues facing them. These speeches are highly varied: to some, it provides an opportunity
to air grievances on speciﬁc issues and conﬂicts, while for others it provides an opportunity to
expound upon a vision for the future of both their own country and the community more broadly.
Though recent work has provided interesting insights into the content (Baturo et al., 2017) of
these speeches, as well as the use of these speeches to derive ideological distances (Gurciullo &
Mikhaylov, 2017), relatively little is known about how these speeches are generated at the level of
politics, and what they can communicate about a country’s broader foreign policy interests. This
despite the fact that especially at the United Nations, these formal, public statements of opinion and
foreign policy interests are both proliﬁc and detailed.

This disconnect between data and theory presents a potential opportunity for both the study of
the United Nations and for the study of diplomacy more broadly. Unlike public threats (Sartori,
2005, 2002; Fearon, 1994), making demands of other countries (Trager, 2017, 2013), or statements
of resolve to deter an adversary (Ramsay, 2011; Sartori, 2002), statements at the United Nations
General Assembly tend to be statements of opinion–speeches delivered to express a particular
country’s view on a variety of issues. Though these statements of opinion may be interpreted in
many ways–such as with respect to communicating resolve–the more fundamental question remains
unanswered of whether simply expressing an opinion at all on an issue communicates anything
noteworthy. This is especially true since at forums like the UN General Assembly, where there is
relatively little constraint on a country’s capability to speak, and almost no constraint on the content
of these speeches.

Statements of opinion often do not fall into the category of a threat, demand, or statement
of resolve; nor is it necessarily intuitive why these statements surface on the issues they do. For

66

example, on 18 December 2002, during the process of voting on the Optional Protocol to the
Convention against Torture, the Kingdom of Bahrain spoke in regards to its vote:

My country’s delegation wishes to explain its position before the vote on the draft
resolution on the implementation of human rights instruments and on the Optional
Protocol to the Convention against Torture and Other Cruel, Inhuman or Degrading
Treatment or Punishment. The Kingdom of Bahrain is making eﬀorts, particularly
within the area of the democratic process, to establish institutions designed to serve the
aspirations of our people, which will help advance the interests of Bahraini citizens.
The Constitution and legislation of Bahrain condemn all forms of discrimination against
or cruel treatment of any human beings within the territory of the Kingdom whether
they are citizens or resident aliens. That is tangible evidence of the great importance
the Kingdom of Bahrain attaches to the preservation of and respect for human rights
in order to promote human dignity and the role of individuals (Kingdom of Bahrain,
A/57/PV.77).

Though this is an otherwise simple and relatively benign political statement, it is somewhat
illustrative, since the statement seems otherwise entirely superﬂuous: the statement occurs in the
wake of a vote which had already taken place–to which Bahrain voted in the aﬃrmative1–which
should have served to communicate Bahrain’s interest in the issue already. Hence, the question:
why would countries state their opinion verbally at an otherwise legally powerless forum like the
UN General Assembly, when actions can serve to communicate these opinions equally as well (or
better)?

In this paper, I argue that the decision to speak on an issue reﬂects some political salience
of that issue, since the country’s leader has made a political calculus that its behavior in the past
has been insuﬃcient to properly communicate an opinion on the issue. Note that this need not
require that a country is stating its "true" opinion: in the example of Bahrain above, whether the

1https://digitallibrary.un.org/record/482408?ln=en

67

Kingdom of Bahrain is actually "making eﬀorts...within the area of the democratic process" and
whether the nation actually "attaches [importance] to the preservation of and respect for human
rights" is less important than the fact that the country’s leadership felt it salient to state an opinion
on the issue of human rights at all. The process is, I argue, a reﬂection of some underlying political
reality–whether or not the content of the statement is actually true in an objective sense. More
concretely, speech represents a political decision to speak about a country’s interests–whether with
the intent to clarify or muddy the waters–to some audience with political relevance to the speaker.
Drawing on an original dataset of all speeches delivered at the UN General Assembly between
1984-2014, I show that (a) countries act in predictable ways with respect to whether they speak
diﬀusely or speciﬁcally across the universe of issues under discussion; and (b) that the decision to
speak on some issues rather than others can be mapped to observable foreign policy behavior–in
this case, co-involvement in civil wars. The ﬁrst provides insight into the political roots of how
and why countries allocate their speech at a forum like the UN General Assembly, given that they
can choose to speak on as many or as few issues as they want. The second shows that the priorities
countries express in their speeches do reﬂect some level of interest, as opposed to being reserved
for either issues countries do not care about, or for exogenously imposed (such as by the UN itself)
issues that are independent of a country’s interests. This paper contributes to a growing body
of literature on diplomacy, showing that countries deliver statements of opinion in part to reﬂect
politically salient issues.

3.1 Diplomacy, communication, and the audiences who listen

A body of literature has steadily grown at the nexus of diplomacy and international politics,
gaining insights into how the things countries say contain–if at all–information about what countries
mean.
In a well-known model of international crises, Fearon (1994) shows that when leaders
"engage the national honor" through threats and demands, they risk potential political costs imposed
by the audiences with the power to remove a leader from power. He shows that this results
in the capacity to communicate credibly, arguing that democracies–beholden directly to voters–

68

can successfully communicate their resolve due to the presence of "strong" audiences. This has
also been expanded to explore autocratic countries by Weeks (2008, 2012), blossoming itself
into its own line of productive literature and data collection eﬀorts (Geddes et al., 2014). This
mechanism of translating audience costs into behavioral diﬀerences among regimes has been
validated experimentally (Tomz, 2007) and has seen extensive theoretical reﬁnement (Schultz,
2001; Tomz & Weeks, 2013; Kertzer & Brutger, 2016) over the paste several decades. Though it
has not been without some skepticism (Snyder & Borghard, 2011; Trachtenberg, 2012) and debate
(Gartzke & Lupu, 2012; Slantchev, 2012; Kurizaki & Whang, 2015), the framework that connects
leaders to the audiences who can impact them is a valuable one with extensive implications for the
study of international politics, even beyond the more speciﬁc mechanism that underlies audience
cost theory.2

This connection between the leaders who speak and the audiences who listen has been leveraged
throughout the study of international diplomacy, with implications for how and whether countries’
statements communicate information. In another landmark work in diplomacy, Sartori (2005) shows
that international audiences can impact the way in which countries communicate. She shows that
through the prospect of reputational costs imposed from international audiences, countries have
incentives to make statements–most notably threats and demands–that reﬂect their actual priorities,
as opposed to being misrepresentations or blustering for the purpose of political maneuvering.
This mechanism adds an international element to the broader question of audiences, and it also
generalizes the speciﬁc question of threats into a more broad study of credibility. It also links the
study of diplomacy with the longstanding question of trust in international politics (Kydd, 2005,
1997; Schweller, 1996; Hobbes & Macpherson, 1968). Much like the work on domestic audiences,
this mechanism has been subject to reﬁnement (Guisinger & Smith, 2002) and debate (Press, 2005),
and highlights the way in which the presence of international–in addition to domestic–audiences
can generate patterns of credible communication.3

2For a useful summary on the speciﬁcity of "audience cost" theory as it relates to the broader

question of communication, see Schultz (2012).

3The role of audiences in communication has also been extended to explain why some statements

69

Diplomatic communication is not limited to the audiences a leader speaks to: it also extends
to audiences a leader does not speak to. This dovetails with a growing literature on how the
presence of opposition groups can facilitate communication (Trager & Vavreck, 2011; Ramsay,
2004; Schultz, 1998); more generally, the mechanisms of communicating via public speech go
beyond the presence of politically relevant audiences, and also include the presence of alternative
audiences to whom a leader chooses not to speak. This is especially relevant to a forum like the UN
General Assembly, where leaders actively eschew some issues in favor of others when broadcasting
their opinions to worldwide audiences. The diversity of issues a leader can speak on, combined
with the issues a leader actually does speak on, creates dimensionality in which incentives can exist
for communication (Trager, 2011).4 Ultimately, I expect that the choice to speak in public, and to
choose to speak on some issues over others, communicates the political salience of those particular
issues in addition to communicating the fact that a politically relevant audience exists for those
issues.

3.1.1 Statements of opinion and the UN General Assembly

These two existing frameworks can be combined to understand speeches at the UN General As-
sembly, with testable implications for patterns of how countries prioritize diﬀerent issues. Given a
possible set of audiences to whom a leader can speak, the informative value of these speeches can
in part come from choosing to target some set over others. In the example of Bahrain above, the
perceived beneﬁt of speaking about Human Rights–even given their aﬃrmative vote–outweighed
the perceived beneﬁt of remaining silent. This could be for a host of reasons: at the time, for
example, the ascension of Hamad bin Isa to king in February of 2002; new human rights initiatives
including steps loosening restrictions on organized labor; and the public designation of Bahrain as
a major U.S. ally (Katzman, 2019). The political decision to speak on the issue of Human Rights
points to both the presence of an audience for these opinions in the ﬁrst place, and the salience of
are delivered in private as opposed to on a public forum (Tarar & Leventoğlu, 2009; Kurizaki, 2007)
4For a slightly diﬀerent take on eschewing some possibilities in favor of others, see Trager (2013)

on how choosing to demand some amount as opposed to all of an issue can convey information.

70

the issue–given that the implemented reforms and aﬃrmative vote was still insuﬃcient to fulﬁll the
political need to express the opinion.5

Note that this does not imply anything about whether, at this time, King Hamad actually
believed in human rights in some objective sense; nor does it imply anything about whether King
Hamad ultimately integrated human rights speciﬁcally into observable policy–since translating
political interests into actual policy requires a complicated process involving additional costs and
considerations (Gartzke & Jo, 2006). However, the decision to speak on the topic does suggest the
presence of a politically relevant audience willing to listen. In the aggregate, while it is diﬃcult
to map any single issue to any single policy outcome, there should be predictable patterns of
issue emphasis within speeches, depending on the characteristics of the audience and the political
characteristics of the speaker.

With the UN General Assembly, the combination of possible audiences and possible issues
becomes especially poignant. With a worldwide membership of 193 countries, and its status as
one of the most prominent public forums in international politics, the set of possible audiences
(including non-state groups, domestic groups, institutional elites, etc.) and the set of possible issues
to discuss are functionally inﬁnite. As a result, allocating time and political messaging toward some
issues rather than others should become a viable method of communicating political priorities.

3.1.1.1 Empirical implications

Given some set of politically relevant interest groups of size , there are two extremes a leader can
employ in a given speech. On one side, a leader will attempt to address multiple interest groups
5This is not necessarily a novel formulation of how the presence of an audience can reveal
information from UN General Assembly speeches. For example, Binder & Heupel (2015, p.242)
argue that "...providing reasons and justiﬁcation for positions in the public sphere is not only an
important political act in itself...but it is also revealing with regard to the beliefs held by the
intended audience. Whether states make statements about the Council’s legitimacy and how these
statements are framed show us what grounds for legitimacy, if any, are believed to resonate within
the community of states" (emphasis added). In this way, the presence of a large, public audience
can be informative with respect to which issues a leader believes will "resonate" most with their
intended, relevant political audience.

71

in a single speech, peppering it with either lip service or genuine concern on all  possible issues.
As noted above, under this framework the prioritization of an issue in a given speech does not
speak to exactly how a leader necessarily feels about a given issue, but rather it speaks to the fact
that there is a politically relevant audience whom the leader beneﬁts from speaking to. This could
be due to political institutions or a large, diverse set of interest groups at home, or it could be
due to a country’s extensive reach into a large variety of issues, and thus a more diverse web of
interconnected issues to juggle at once. Contrast this with the other extreme, where a leader will
only devote speech to a single issue  ∈ , to either focus maximum impact on a single politically
powerful interest group, or to focus limited political capacity onto a particular, relevant issue area.
From this, we should see certain observable implications. A leader with a diverse set of
interest groups to whom they are politically beholden should, on average, lean towards the ﬁrst
extreme–where they tend to speak on more issues in any given speech in order to placate multiple
audiences. This should most notably be visible in political institutions: democratically elected
leaders should, all else equal, be beholden to a larger number of politically relevant audiences
than Autocratic countries, drawing more beneﬁt from placating opposition parties and voters
with diﬀerent priorities. This in contrast to Autocratic regimes, which as Li (2006) puts it, "are
characterized by narrow elite control and a small winning coalition".6 Given these characteristics
we should expect that, in any given speech–delivered on a speciﬁc issue, at a speciﬁc time, and in
a speciﬁc context–Democratic countries should show a tendency to placate multiple audiences at
any time in any given speech, resulting in a more diﬀuse focus within any speech.

In addition to the size of the politically relevant audience, a country should show a propensity
towards diﬀuse speech when it is more capable of actually acting on a larger number of issues.
These should tend to be more powerful countries, with both the economic and military power
to actually sustain a large, diverse portfolio of issues in which they are involved. Conversely,
countries with less capacity to actually have tentacles in multiple issues should, on average, speak
more speciﬁcally, focusing in on the issues where they have inﬂuence, or speaking only to a narrow

6Also see Lektzian & Souva (2007, 857)

72

set of issues where the risk of oﬀending strong political opposition can be minimized.7 Much like
democratic leaders, countries with more capacity to act will tend to have a larger set of politically
relevant interest groups to whom they beneﬁt from speaking to, and thus should be more diﬀuse in
the speeches they deliver.

H1: Countries with a larger set of politically inﬂuential audiences at home will speak
more diﬀusely on average

H2: Countries with the capacity for involvement in more issues should speak more
diﬀusely on average

3.2 Data and issue composition

To test these claims, I draw from the text of UN General Assembly speeches delivered between
1984-2014 (Mahmood, 2020b).8 This corpus is distinct from the one collected by Baturo et al.
(2017), which focuses primarily on yearly speeches delivered under the General Debate agenda
item; the corpus I use in this paper includes a more comprehensive set across all plenary sessions
of the UNGA.9 I scrape, clean, and parse the text primarily in Python.

The concepts of diﬀuseness and speciﬁcity in speech remain yet to be operationalized. There
are potentially many methods of doing so. For example, one option would rely on the UN’s own
internal measurement: in this case, using agenda items. For example, during session 57/plenary
session 47, the UN General Assembly passed three resolutions, each on diﬀerent topics: in this
case, resolutions A/RES/57/10; A/RES/57/9; and A/RES/57/8. Respectively, these covered agenda
7In prior work, the author found that weaker countries spoke much less than stronger countries
overall at the UN General Assembly. This follows from a similar framework:
the cost-beneﬁt
analysis of speaking on issues and either risking saying the incorrect opinion or failing to address
an important counterpoint should be far riskier for weaker countries or leaders, who have little
capacity to even set an agenda on issues regardless.

8Data collection for the years 1945-1984 is ongoing
9All results presented here remain substantively similar when applied to the UN General Debate

corpus from Baturo et al. (2017).

73

item 34–on Bosnia & Herzegovina; agenda item 14–on a report of the IAEA; and agenda items
21 (d) and 37–on a war-torn Afghanistan. By the logic outlined above, we could deﬁne speciﬁcity
by the variety of agenda items a country speaks on: in this case, we would expect democratic and
powerful countries to speak on a greater number of diﬀerent topics than average.

While this is a promising method of deﬁning the space of potential topics a country can speak on,
it comes with a prohibitively large amount of work to gather the requisite data: within each plenary
session, as the example of A/57/PV.47 demonstrates, there can be multiple agenda items, each with
associated speeches. To do this manually would require copying and pasting text from potentially
hundreds of transcripts, requiring extensive resources. Automating this is a more scalable solution,
but is not without its diﬃculties: Figure 3.1 provides a snapshot of the text of the UN’s oﬃcial
transcript of A/57/PV.47, with a particular focus on the transition between agenda items. The
translation of this to raw text removes the bold-face formatting, and the syntax of when an agenda
item starts and ﬁnishes is not consistent across transcripts. Thus, I turn to alternative methods to
measure the diﬀuseness and speciﬁcity of topics.

The raw text of speeches contains a rich level of detail about the topics a country can speak
about. This level of detail trancends even the stated agenda item of the plenary session:
for
example, during a speech by India’s Hardeep Singh in A/57/PV.47–delivered under the agenda item
of the IAEA, and partially visible in Figure 3.1–the topics of the IAEA as an organization; Nuclear
power; and global climate change were all raised during the course of three diﬀerent paragraphs.
Thus, in addition to overcoming the diﬃculties with separating transcripts by their respective
agenda items, drawing from the content of the speeches themselves allows the opportunity to more
comprehensively understand how countries speak on some topics with respect to the population of
topics across all speeches at the UN General Assembly.

To measure this, I turn to methods of natural language processing, and use a Structural Topic
Model (Roberts et al., 2014) to measure the composition of individual speeches. This method
represents each speech as a K-dimensional vector, where K is some number of possible topics.
For example, if the above speech by India were composed of three equal-length paragraphs, each

74

Figure 3.1: Snapshot of UN General Assembly transcript

on a speciﬁc topic, in theory a 3-topic STM could represent this speech as the vector [0.33, 0.33,
0.33], where the speech is shown to contain 33% attention to each topic.10. I use the algorithm
proposed by Mimno & Lee (2014) to select a number of topics (Roberts et al., 2014, p.12), which
yields 81 total topics.11 I do not include covariates in the model, meaning the model is reduced to
a logistic-normal topic model, or a correlated topic model (Blei & Laﬀerty, 2006).

10Of course, this is a simpliﬁed example, since statistically there is no way to guarantee speciﬁc
topics arise from a given text. However, using this method is still valuable, since it provides the
best ﬁtting model of the text–as represented by individual words and whether or not they occur in
the same document–into a mixture of K latent, unobserved topics

11As Grimmer & Stewart (2013) correctly point out, there is no truly "correct" answer to the
number of real topics present in the data. While this approach leans heavily on automated methods,
I also include in later sections a qualitative approach as well by aggregating the topical composition
into a set of qualitatively informed groups.
I detail each group in the appendix; however, for
example, this method aggregates two similar topics such as Humanitarian assistance in natural
disasters and Humanitarian coordination and assistance for emergency responses into a single
category, Humanitarian Assistance. In part, the categories are informed by dkomain knowledge
gained from extensive reading of these speeches, as well as previously estimated topic models
on much smaller (and arbitrary) numbers of topics such as 15, 20, 25, and 30 topic models. To
aggregate the composition, I simply add the estimated proportions of relevant topics among each
category; in total, my categorization yielded 22 categories across the 81 estimated topics–including
one topic simply consisting of document-level noise. I re-estimate all models on this 22-category
composition, yielding substantively similar results.

75

3.2.1

Issue diﬀuseness and speciﬁcity

The primary hypotheses above make predictions about the diﬀuseness and speciﬁcity of any given
speech. More concretely, given that we can represent any given speech as some k-dimensional
vector composed of  issues,

 = [ 1, 2, 3, ..., ]

(3.1)

where the vector  sums to 1, we can represent this concept as the maximum value of any given
value  within . The sum constraint allows for this, since each value of  is dependent on the
others, and thus a higher maximum value implies that the others will be lower. A more speciﬁc
speech, then, will have a higher maximum, where the speaker is delivering a speech emphasizing
some topic ∗ disproportionately relative to the other topics. In contrast, a diﬀuse speech will have
a lower maximum, where the speaker is delivering a speech emphasizing some topic ∗ relatively
similarly to the others. Corresponding to the proposed theory, then, each value of [ ] corresponds
roughly to a given political audience who is listening for that given issue.12

Each speech, however, exists at the document (i.e.

the speech) level, while the covariates
measuring power and democracy are at the yearly level. Thus, we are left with several possibilities.
The ﬁrst possibility is to simply estimate the STM model across all words spoken in a given year.
This is problematic, though, since we lose the granular speech-level information that the speeches
provide. The second possibility is to model each speech as an independent observation, repeating
the yearly covariate for every speech. This would likely over-estimate the certainty in the results,
since the covariates are not measured at each point of observation.

The third option, which I employ here, is to aggregate the individual speech compositions over
the year. Given that each speech is a diﬀerent length, it would not be prudent to simply average
each topical proportion  for a given country in a given year, since it would over-emphasize
lower-length speeches and under-emphasize higher-length speeches. Instead, I distribute the word
12This is, no doubt, a highly simpliﬁed and stylized version of this mechanism, since a given
audience and interest group can have multiple interests simultaneously–albeit usually correlated
ones. The speciﬁcity and diﬀuseness of the interest groups themselves is an interesting area for
future research, especially in how it links to the content of political speech.

76

count  for each speech  across its respective composition, creating an estimated word count per
topic , per speech

∗
, =  ∗ ,

(3.2)

Next, I calculate an estimated yearly word count  by summing all estimated word counts

across speeches  ∈ , for all topics 


=1


 =

∗
, ∀ 

(3.3)

I do this for all countries, for each year, resulting in a country-year dataset of estimated words

spoken on each of the 81 topics.

The question of interest, though, is on a country’s emphasis across topics, rather than which
countries speak more than others. Thus, I revert this vector of estimated words spoken  =
[1, 2, 3, ..., ] back to a compositional variable ∗ in order to capture this emphasis by
dividing all estimated words spoken by a given country  by the total number of words spoken in
that year, reverting back to a vector

∗ = [ ∗

1, ∗

2, ∗

3, ..., ∗
]

(3.4)

Finally, each vector ∗ exists on the K-1 simplex, meaning that the maximum value is constrained
to be less than 1. To transform the vector to euclidean space, I perform an Aitchison Log Ratio
transformation on the compositional variable  Aitchison (1986), deﬁned as:

(cid:34)

(cid:33)

(cid:32) ∗

1
∗


(cid:33)

(cid:32) ∗

2
∗


(cid:33)(cid:35)

(cid:32) ∗

−1
∗


(∗) =


; ...; 

; ...; 

(3.5)

deﬁning ∗

 by choosing one topic out of the 81 as the reference topic. Finally, one beneﬁt of
Topic Models is the ability to qualitatively understand the content of each topic by analyzing the
most probable words given any topic. This allows for additional cleaning, which is important in this
case since the goal is to understand issue speciﬁcity. Some topics seem to be primarily referring
to grouped statements, where for many countries part of their speech is devoted to diplomatic
platitudes "on behalf of" their respective foreign policy group. In this case, the three groups present

77

in the topic models are the European Union; Non-Aligned Movement; and Carribean Community.13
There is also a topic devoted purely to noise due to parsing the UN text, with the top words referring
to the Dag Hammarskjold Digital Library, where the transcripts are housed. For the analysis, I
remove these topics and calculate the maximum value in (∗), omitting these four topics.14

The dependent variable of interest–a measure of issue speciﬁcity at a given country-year level
of analysis–is a continuous measure of the maximum value in (∗). Figure 3.2 provides a kernal
density plot of the maximum ALR values aggregated across all countries and all years. The data
is generally normally distributed; to model the data, I estimate an Ordinary Least Squares model.
Results are substantively similar when estimating a Robust Linear Model instead–which down-
weights inﬂuential observation according to an M-estimator, using iterated-weighted-least-squared
regression (IWLS) (Venables & Ripley, 2013).15

3.2.2 Model and variables

The primary independent variables of interest are a country’s capacity to inﬂuence, and the political
relevance of diverse audiences. I use three diﬀerent measures to account for the ﬁrst variable. First,
I include the Composite Indicator of National Capabilities (Singer et al., 1972, version 5), which
provides as a continuous variable an indexed score aggregating military, population, economic, and
other measures of a country’s general level of power. Second, I include a measure for a country’s
Gross Domestic Product Per Capita (Food & of the United Nations, 2019b) to capture economic
13Note that these are not simply regional topics. For example, the European Union topic is
not simply a topic about European issues, but rather includes words like "European", "Union",
"Welcome", "Committee", "Particular", pointing more to an administrative and diplomatic topic as
opposed to a more substantive one.

14I also estimate all models with these included. The results are substantively similar, though
as expected the regional ﬁxed eﬀects change slightly, with Europe showing a tendency to be more
speciﬁc on average, and France/the UK being more diﬀuse than average–more on par with the
United States. Key ﬁndings, however, do not change.

15For robustness, I also estimate all models by excluding any observation with a Dependent
variable less than 6, a threshold that came from visual inspection of a qq-plot. I also estimate all
models (OLS, RLM, and threshold corrected) using bootstrapped standard errors via the sandwich
package in R (Zeileis, 2004, 2006). The results are substantively similar to those presented here.

78

Figure 3.2: Distribution of maximum ALR value of year-aggregated speeches, 1984-2014

power in particular. Finally, I include a measure of Nuclear capability for similar reasons, since it
serves as a uniquely stark measure of a country’s coercive leverage–especially in the post-World War
2 era. For this, I use Fuhrmann & Tkach’s (2015) data on Nuclear Latency to create a 4-category
variable composed of None, Either lab or pilot, Both lab and pilot, and Full.16

For the second, I use the Uniﬁed Democracy Scores (Pemstein et al., 2010), which provide a
latent variable estimate for democracy across ten existing scales, including among others the Polity
IV score (Marshall & Jaggers, 2018); the Freedomhouse score (House, 2014); and the (Cheibub
et al., 2010) measures.17
I also include a measure of a country’s population (Food & of the
United Nations, 2019a) as a second measure for the diversity of a country’s domestic constituents.
16Including three diﬀerent variables allows for a more robust view of the hypothesis, since there
are considerable tradeoﬀs with using any single one. As the analyses show, for example, while
CINC score is a more generalized score for countries, it is highly correlated with its constituent
parts–in some cases like population, up to around 95%–rendering analyses with the constituent
parts included at risk for statistical artifacts. A measure like GDP Per Capita allows us to separate
between the constituent parts, but it also focuses in more on economic power, and also results in
substantially more missing observations. Finally, Nuclear Latency provides a stark, clear measure
of coercive power with relatively good coverage, but it sacriﬁces the granularity of a continuous
measure like the CINC score.

17Full list, detailed citations, and information available online. See Pemstein et al. (2010).

79

0.000.050.100.150.20051015maxALRdensityMax ALR distributionOften, UN General Assembly speeches are delivered via grouped statements. Despite the
otherwise independent nature of speeches, this particular phenomenon must be accounted for; thus,
I include a binary variable for whether or not a country is a member of the EU, Non-Aligned
Movement, African Union, or ASEAN, as well as a variable for whether they are a member
of no group at all. The reference category, then, is any other group.18 To further account for
interdependence between countries, I also divide Bailey et al.’s (2015) UN General Assembly
Vote-based ideal point measure into 5 mutually exclusive categories, and include a categorical
indicator of which category a country belongs to.

From the example of Bahrain above, we may expect that a country’s voting behavior in a given
year will inﬂuence how they speak: given the relatively rare incidence of dissident voting at the
UN General Assembly (i.e. "no" votes and "abstain" votes), countries with higher proportions
of non-aﬃrmative votes may have diﬀerent tendencies to speak diﬀusely or speciﬁcally in their
speech. Thus, I include a logit-transformed19 proportion of non-yes votes per country out of the
total number of resolutions in a given year.

Finally, I also include the total number of words a country has spoken across topics, a measure
of a country’s human rights (Fariss, 2014), a count of Inter-Governmental Organizations to which
the state is party in a given year (Pevehouse et al., 2004), the Region to which a country belongs20, a
mutually exclusive categorical variable for the ﬁve major powers (including a category for non-P5),
and year-ﬁxed eﬀects to account for the overall level of speciﬁcity present among the international
community in a given year.
I log-transform highly skewed variables in the analysis, including
population; CINC score; and GDP Per Capita.

18There are a total of 43 groups identiﬁed, on behalf of which speeches are commonly delivered
19That is, given some proportion  = ¬/ non-yes votes ¬ on  resolutions, the logit-
transformed proportion ∗ would be calculated as ∗ = (( /(1 − )); for those with zero
non-yes votes, I substitute 0 for = 0 with  = 0.0000001.

20Region, in this case, consists of Americas; Asia/Oceania; Europe; Middle East and North

Africa; and Sub-Saharan Africa.

80

Figure 3.3: Descriptive relationship between measures of power and issue speciﬁcity

3.3 Analysis

Table 3.1 provides a correlation matrix between the main variables of interest. Figure 3.3
provides a Loess-smoothed bivariate relationship between three measures of power and a log-
transformed maximum-ALR value, and Figure 3.4 provides the same for the two measures of
political audience size. The descriptive patterns are consistent with the expected patterns; while
GDP per capita does show some curvilinearity, this is largely driven by relatively few lower-income
countries. However, to be thorough, I also include a quadratic term for power in some model
speciﬁcations; the results are not substantively aﬀected.

log_maxALR
maxALR
logcinc
loggdppcFAO
NucLat_Numeric
UDSmedian
logtpopFAO
centIGOmem

log_maxALR maxALR logcinc
-0.11
-0.15
1.00
0.16
0.56
0.01
0.93
0.58

0.89
1.00
-0.15
-0.02
-0.13
0.07
-0.10
0.06

1.00
0.89
-0.11
0.00
-0.10
0.07
-0.07
0.08

loggdppcFAO NucLat_Numeric UDSmedian logtpopFAO centIGOmem
0.08
0.06
0.58
0.47
0.36
0.51
0.52
1.00

-0.07
-0.10
0.93
-0.11
0.53
-0.08
1.00
0.52

0.00
-0.02
0.16
1.00
0.19
0.63
-0.11
0.47

-0.10
-0.13
0.56
0.19
1.00
0.12
0.53
0.36

0.07
0.07
0.01
0.63
0.12
1.00
-0.08
0.51

Table 3.1: Variable correlations

Table 3.2 provides output from the Ordinary Least Squares model, across diﬀerent model spec-
iﬁcations to account for the high correlation between CINC score and its constituent parts within

81

1.41.51.61.7−12−8−4logcinclog_maxALRlog CINC vs log max ALRBivariate Loess smooth1.451.501.551.604681012loggdppclog_maxALRlog GDPpc vs log max ALRBivariate Loess smoothlllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll0.81.21.62.02.40None1LabOrPilot2LabAndPilot3FullNucLatlog_maxALRNuclear Latency vs log max ALRBivariate boxplotFigure 3.4: Descriptive relationship between audience size and issue speciﬁcity

the model (namely population and GDP per capita). Support for hypothesis 1 is consistent across
speciﬁcations: as countries become more democratic, and as they become more populated, the
composition of speeches become much more diﬀuse. This is consistent with the theory that as
countries have more politically relevant audiences to whom they are beholden, their speech com-
position will reﬂect that diversity. More importantly, it casts doubt on the claim that emphasizing
particular issues is about pleasing some exogenous audience (for example, the UN’s administration,
an elite group of interested NGOs, or even a single world superpower), where we might expect little
variation across regime types. Figure 3.5 plots the predicted linear relationship between Democracy
and speech speciﬁcity, and Figure 3.6 plots the same for population.21

The results for powerful countries are mixed, but generally consistent.

In both the case of
Nuclear capacity and GDP per capita, countries with greater capacity tend to speak more diﬀusely,
as expected. However, the CINC score is ﬂipped in a model that includes population; given the very
high correlation of around 94% between CINC score and population, it is possible this is a statistical
21The results are substantively similar when we use Polity score (Marshall & Jaggers, 2018)
instead of UDS; it also reveals an interesting pattern where the high speciﬁcity is mostly concentrated
in the most autocratic countries–with scores of -9 and -10.

82

1.51.61.71.81.9−2−1012UDSmedianlog_maxALRUDS score (median) vs log max ALRBivariate Loess smooth1.41.51.61.71.8912151821logtpopFAOlog_maxALRlog population vs log max ALRBivariate Loess smoothNucLat3 - Full

cent IGO mem

log(CINC)

log(Pop)

g_AU

g_EU

cent Word Count

3.33∗∗∗
(0.28)

NucLat1 - LabOrPilot

NucLat2 - LabAndPilot

(Intercept)

log(GDPpc)

Democracy

Hum Rights

−0.25∗∗∗
(0.02)
0.36∗∗∗
(0.11)
0.10
(0.10)
0.16
(0.13)
−0.16∗∗∗
(0.03)
−0.03∗
(0.02)
−0.07∗
(0.03)
−0.06
(0.05)

6.87∗∗∗
(0.46)
−0.08∗∗∗
(0.02)
−0.24∗∗∗
(0.02)
0.40∗∗∗
(0.10)
0.20∗
(0.09)
0.18
(0.12)
−0.14∗∗∗
(0.03)
−0.04∗
(0.02)
0.00
(0.04)

All Vars CINC (no parts) Excl CINC Excl Pop
8.56∗∗∗
3.59∗∗∗
(0.30)
(0.99)
−0.04∗
−0.10∗∗∗
(0.02)
(0.02)
−0.24∗∗∗
−0.25∗∗∗
(0.02)
(0.02)
0.39∗∗∗
0.40∗∗∗
(0.11)
(0.10)
0.20∗
0.15
(0.09)
(0.09)
0.16
0.19
(0.12)
(0.12)
−0.17∗∗∗
−0.12∗∗∗
(0.03)
(0.03)
−0.04∗
−0.01
(0.02)
(0.02)
−0.05
0.00
(0.03)
(0.04)
−0.09∗∗∗
0.06
(0.03)
(0.01)
−0.20∗∗∗
(0.04)
−0.23∗
(0.11)
0.37∗∗∗
(0.08)
0.15∗∗
(0.05)
−0.10
(0.09)
0.29
(0.28)
0.32
(0.23)
0.20
(0.28)
0.26
(0.28)
majpow - USA
0.55
(0.29)
logit_Prop_NoVotes_AbstAsNo −0.02∗∗∗
(0.00)

−0.14∗∗∗
(0.02)
−0.28∗∗
(0.11)
0.37∗∗∗
(0.08)
0.16∗∗
(0.05)
−0.09
(0.09)
0.30
(0.28)
0.37
(0.23)
0.25
(0.28)
0.29
(0.28)
0.62∗
(0.28)
−0.02∗∗∗
(0.00)

−0.31∗∗
(0.11)
0.42∗∗∗
(0.08)
0.17∗∗∗
(0.05)
−0.10
(0.09)
0.43
(0.28)
0.54∗
(0.23)
0.39
(0.28)
0.40
(0.28)
0.69∗
(0.28)
−0.02∗∗∗
(0.00)

g_NAM

g_ASEAN

majpow - RUS

majpow - UK

majpow - FRA

majpow - Other

−0.24∗
(0.11)
0.42∗∗∗
(0.08)
0.18∗∗∗
(0.05)
−0.15
(0.08)
0.47
(0.29)
0.57∗
(0.24)
0.38
(0.28)
0.42
(0.29)
0.65∗
(0.29)
−0.02∗∗
(0.00)
0.00
(0.00)
0.02
(0.15)
Yes
Yes
Yes
0.82
0.82
4408

log(CINC)2

g_NONE

Year FE
Region FE
UNGA Ideal Pt Cat
R2
Adj. R2
Num. obs.
∗∗∗  < 0.001; ∗∗  < 0.01; ∗  < 0.05

Yes
Yes
Yes
0.82
0.82
4351

Yes
Yes
Yes
0.82
0.82
4351

Yes
Yes
Yes
0.82
0.82
4351

Table 3.2: OLS Regressions

83

Figure 3.5: Predicted max ALR value across Democracy Score

Figure 3.6: Predicted max ALR value across log-Population

84

6.97.27.57.8−2−1012UDSmedianmaxALR6.57.07.58.08.5912151821logtpopFAOmaxALR(Intercept)

cent Word Count

NucLat1 - LabOrPilot

NucLat2 - LabAndPilot

NucLat3 - Full

Democracy

Human Rights

cent IGO mem

log(CINC)

Milex
7.99∗∗∗
(0.70)
−0.22∗∗∗
(0.02)
0.33∗∗
(0.10)
0.04
(0.09)
−0.24
(0.13)
−0.10∗∗
(0.03)
−0.03
(0.02)
−0.04
(0.04)

g_zmNONE

majpow - RUS

majpow - UK

g_NAM

g_ASEAN

majpow - FRA

g_AU

g_EU

majpow - Other

−0.23∗
(0.11)
0.29∗∗∗
(0.07)
0.15∗∗
(0.05)
−0.02
(0.09)
0.31
(0.24)
0.28
(0.27)
0.08
(0.23)
0.09
(0.27)
0.22
(0.27)
majpow - USA
0.14
(0.29)
logit_Prop_NoVotes_AbstAsNo −0.02∗∗
(0.01)
−0.07∗
(0.03)
−0.22∗∗
(0.08)
0.01∗∗
(0.00)
−0.11∗∗∗
(0.03)
YES
YES
YES
0.84
0.83
3775

Year FE
UNGA IP FE
Region FE
R2
Adj. R2
Num. obs.
∗∗∗  < 0.001; ∗∗  < 0.01; ∗  < 0.05

log(GDPpc)

log(Milex)

log(Milex)2

log(Pop)

Table 3.3: OLS Regression - Military Expenditure

artifact. Table 3.3 provides output from a similar model, using military expenditure instead of
CINC score, retaining population and GDP in the model. While this model has substantially fewer
observations due to missing data, the coeﬃcient is in the expected direction. Given this result
combined with the descriptive data, prior ﬁndings on overall UN engagement22, and results from
alternative measurements, there does seem to be general–though not conclusive–support for the
hypothesis that more powerful countries will speak more diﬀusely across issues.

22Also note that higher word count, which was shown in previous work to be strongly related to

a country’s capacity, is associated with more diﬀuse speech.

85

Figure 3.7: Snapshot of UN General Assembly Resolution metadata

3.3.1 Alternative measures and the content of speeches

3.3.1.1 Resolution topics

To further validate the expected patterns, I turn to an alternative measure for the number of issues
a country decides to speak on, based more on the UN General Assembly’s own coding of issue
areas. While parsing agenda items across transcripts poses a large challenge, UN General Assembly
Resolutions are also associated with agenda items, providing an opportunity to at least approximate
agenda items. For this, I scrape the UN’s website for metadata on all UN General Assembly
resolutions between 1984-2014.23 I then draw from voting data collected by Erik Voeten (et al)
(Bailey et al., 2015) to match individual plenary sessions per resolution (identiﬁed in the voting
data, going back only through 198424) with the metadata scraped from the web. Figure 3.7 provides
a screenshot of this table from the UN’s website for session 67.

Finally, I keep only the subset of speeches delivered during these particular plenary sessions
which were associated with UN General Assembly votes. This provides a rough approximation
for the number of unique agenda items where a country spoke in a given year: for example, if a
given plenary session saw votes on agenda items 121, 54, 19, and 13, a country that spoke during
that plenary session can be approximately said to have spoken on four agenda items.25 Thus, we
can approximate the number of agenda items a country spoke on in any given year in order to
23I use the URL structure https://www.un.org/depts/dhl/resguide/r54esolutionsableng.htm,
replacing the number 54 with all sessions from number 35 through 67. For sessions above 67, I use
the URL structure https://www.un.org/depts/dhl/resguide/r69esolutionsablen.htm, replacing the
number 69 with all sessions from 68 through 69.

24Email correspondence with Erik Voeten
25Parsing exactly which resolution the country spoke on reverts back to the original problem in

the transcripts exempliﬁed by Figure 3.1.

86

further test the hypotheses posed here. More speciﬁcally, we can model the proportion of estimated
agenda items that a country spoke on, out of the possible agenda items discussed in a given year.
We should expect that as the number of politically inﬂuential audiences increases, and the capacity
to act in a diverse number of issues rises, so too should the coverage of speech across the possible
agenda items under discussion by the UNGA. We can make a similar dependent variable for the
proportion of Plenary Sessions (out of the possible number of Plenary Sessions where a Resolution
was voted upon), and the proportion of Resolutions on which a country spoke.

Given that the dependent variable here is a true proportion, and that the measure of the variable
does not allow for the granularity of binary successes and failures, I model the proportion using a
beta distribution, estimating a beta regression in R (Zeileis et al., 2016). Table 3.4 provides output
from 6 models; for each of the three dependent variables, I estimate a model including and excluding
Population as a covariate, given the concerns listed above regarding the strong correlation between
CINC score and population. The results remain generally consistent with the expectations given
above. The only exception is the Nuclear capabilities measure; while the distinction between low-
level and no nuclear capability is positive and signiﬁcant, the remaining categories are statistically
indistinguishable from the no-capability category. While this may suggest curvilinearity, additional
tests on CINC score and GDP per capita do not show any curvilinearity, and the coeﬃcient estimates
from the ﬁve major powers are all substantially higher than the non-P5 category, casting doubt on the
claim that the most powerful countries speak on fewer issues than their middle power counterparts.
Overall, there is general support for the claims that both the broadness of politically relevant
audiences and the capacity for action across issues is associated with a greater diversity of issues
on which a country will speak.

3.3.1.2 Dirichlet Regression

The maximum ALR method of modeling speciﬁcity and diﬀuseness allows for the use of more
conventional modeling techniques, but it sacriﬁces the richness of the information estimated by
the structural topic model. From this model, recall that each speech can be represented as some

87

(Intercept)

loggdppcFAO

wc

NucLat1LabOrPilot

NucLat2LabAndPilot

NucLat3Full

UDSmedian

HRfariss

logcinc

logtpopFAO

g_AU

g_EU

g_NAM

g_ASEAN

logIGOmem

majpowLabFRA

majpowLabOther

majpowLabRUS

majpowLabUK

majpowLabUSA

Year Fixed Eﬀ
UNGA ideal pt
World Regions
Pseudo R2
Log Likelihood
Num. obs.
∗∗∗  < 0.001, ∗∗  < 0.01, ∗  < 0.05

PV
−2.99∗∗∗
(0.33)
0.05∗∗∗
(0.01)
0.00∗∗∗
(0.00)
0.05
(0.04)
0.04
(0.03)
−0.07
(0.05)
0.10∗∗∗
(0.01)
−0.00
(0.01)
0.04∗∗∗
(0.01)
0.03∗
(0.01)
0.10∗∗
(0.04)
0.02
(0.03)
−0.02
(0.02)
0.00
(0.03)
0.09∗∗∗
(0.03)
−0.22∗
(0.10)
−0.33∗∗∗
(0.08)
−0.08
(0.09)
−0.24∗
(0.10)
−0.19
(0.11)
Yes
Yes
Yes
0.78

8986.94
4965

PV_NoPop
−2.39∗∗∗
(0.16)
0.05∗∗∗
(0.01)
0.00∗∗∗
(0.00)
0.05
(0.04)
0.05
(0.03)
−0.07
(0.05)
0.11∗∗∗
(0.01)
−0.01
(0.01)
0.06∗∗∗
(0.00)

0.11∗∗
(0.03)
0.01
(0.03)
−0.02
(0.02)
0.00
(0.03)
0.11∗∗∗
(0.03)
−0.23∗
(0.10)
−0.36∗∗∗
(0.08)
−0.10
(0.09)
−0.25∗∗
(0.10)
−0.21
(0.11)
Yes
Yes
Yes
0.78

8983.78
4965

RES
−3.69∗∗∗
(0.64)
0.09∗∗∗
(0.01)
0.00∗∗∗
(0.00)
0.15∗
(0.07)
0.01
(0.05)
0.06
(0.09)
0.15∗∗∗
(0.02)
−0.03∗
(0.01)
0.02
(0.02)
0.07∗∗
(0.03)
0.19∗∗
(0.07)
0.14∗∗
(0.05)
−0.01
(0.04)
0.08
(0.06)
0.01
(0.05)
−0.06
(0.16)
−0.16
(0.15)
−0.14
(0.18)
−0.09
(0.16)
0.35
(0.18)
Yes
Yes
Yes
0.59

7013.49
4965

RES_NoPop
−2.22∗∗∗
(0.31)
0.07∗∗∗
(0.01)
0.00∗∗∗
(0.00)
0.16∗
(0.07)
0.03
(0.05)
0.03
(0.09)
0.17∗∗∗
(0.02)
−0.04∗∗∗
(0.01)
0.07∗∗∗
(0.01)

0.20∗∗
(0.07)
0.13∗∗
(0.05)
−0.02
(0.04)
0.08
(0.06)
0.05
(0.05)
−0.08
(0.16)
−0.24
(0.15)
−0.20
(0.17)
−0.12
(0.15)
0.32
(0.18)
Yes
Yes
Yes
0.59

7008.74
4965

AG
−3.68∗∗∗
(0.64)
0.09∗∗∗
(0.01)
0.00∗∗∗
(0.00)
0.14∗
(0.07)
0.00
(0.05)
0.04
(0.09)
0.15∗∗∗
(0.02)
−0.03∗
(0.01)
0.02
(0.02)
0.07∗∗
(0.03)
0.18∗∗
(0.07)
0.13∗∗
(0.05)
−0.01
(0.04)
0.07
(0.06)
0.01
(0.05)
−0.07
(0.16)
−0.18
(0.15)
−0.15
(0.18)
−0.10
(0.15)
0.29
(0.18)

AG_NoPop
−2.21∗∗∗
(0.31)
0.07∗∗∗
(0.01)
0.00∗∗∗
(0.00)
0.15∗
(0.07)
0.02
(0.05)
0.02
(0.08)
0.17∗∗∗
(0.02)
−0.04∗∗∗
(0.01)
0.07∗∗∗
(0.01)

0.19∗∗
(0.07)
0.12∗
(0.05)
−0.02
(0.04)
0.07
(0.06)
0.05
(0.05)
−0.09
(0.16)
−0.25
(0.14)
−0.21
(0.17)
−0.13
(0.15)
0.26
(0.18)

0.59

6996.32
4965

0.59

6991.71
4965

Table 3.4: Beta regression results, proportion of Plenary meetings; Resolutions; or Agenda items
connected to a country speaking (precision model results omitted)

88

composition of  topics, represented by a vector  = [ 1, 2, 3, ..., ], where the components of
 sum to 1. We can model this directly as a Dirichlet distribution, which is a generalization of the
Beta distribution across multiple dimensions  on the  − 1 simplex. The distribution models both
expected value and precision explicitly; Maier (2014, p.6) derives expected values and likelihood
functions for the distribution, requiring only a reference category speciﬁed by the user prior to
estimating the regression.

For this, I collapse the 81-topic structural topic model into 20 topics, each of which were
qualitatively speciﬁed post-hoc following estimation of the STM. This provides a more concrete
interpretation of similar topics, and also provides an alternative method of determining whether
speciﬁcity and diﬀuseness of speech still follows from diﬀerently speciﬁed topics. As noted brieﬂy
earlier, this method would aggregate similar topics such as Humanitarian assistance in natural
disasters and Humanitarian coordination and assistance for emergency responses into a single
category, Humanitarian Assistance; I conduct this aggregation at the level of individual speeches
by adding together the component topics within each topic, retaining the compositional character
of the now 20-simplex. Table 3.5, Table 3.6, Table 3.7, and Table 3.8 provide tables of these 20
categories (split across tables for readability), with their constitutive topics from the 81 topic model
listed as well.

89

DiplomacyGroup
GroupMemBehalfWorkHonorThnk
EurUnionWelcomeCommitParticul
DevNamPeacekeep
UngaNamResolutWork

Development
DevCntryTradeIndig
DevSocEconSummit
DevCoopHighWillEﬀort
DevCntryDebtNegot
DevPeopYouthGovernSudanSomalia CaricomCommunDevelopSmallReg
DevClimChngSustainSmallGlob
DevPovertPerCentAchievEduc
DevEconTradeInternCntryProduct
DevProgramGovNatSector
IntDuwaitPeaceSwitzDevelop

Aspirational
IntWorldPeace
NationsMustIntPeaceSec
WillProcessYearHope
WorldNationHumanPeacRight
OrgNationTwelvNewWillWorld
WorldChallengDevelopWill
StrongLetStandTogBetterAgend
WorldMustWillCanOne
WorldPeacWarMustHuman
IntMustOneMakeProbPossib
RemindEndurPursLosePartConstant DevEconPeaceSocialOrgan
DevSocMigratPoliciPovert
FinancCrisisGroupNatWorkSituat
DevCntryImplemGoalCommit
DevAfricaSupportContinPartner

Governance
ApartheidRacistRegim
IntDemGovElectSuppSecur
ApartheidRegimGovernSanction CommPeacebuildActivReportFund

HumAsst
HumanitAsstNaturDisastDev
HumanitCoordAsstNationEmergResp

Table 3.5: Collapsed Topics (1-5)

90

HumRights
HumrightsProtectUnivDecl
SlavDiscrimRaciHumTrade

Independence
IndependColoniResolut DelegRepresCharterStatement

Institution

UNSCReformWOrk
ProcMeetHighlevelPlenaryReport
GroupProposProgramSecGenBudg
CommItemRepAgDoc
ForumFrancNatSmallRegMemb
SecGenOrganOperSupportReform
SecCouncilMemReportWorkIss
GenAssembSesWorkReportResolut
DraftResAdoptVoteConsensusPar

IntCoop
IntCoopOrganIslamDevelop

IntLaw
IntCourtLawJustice
IntLawSeaFish
IntLawConventComissSeaLegal
IntTribunCrimeRwandCoopGenocid
IntCovnventMineTraﬃck

Table 3.6: Collapsed Topics (6-10)

91

IntPeaceSec
IntCoopOrganPeacSecur
RefugDisplSportPeac
PeacConfSolutSettlementNegot
ConfSecurPreventCoopOSCregion

Noise
DigitLibDagHammarsk NucAgenIAEAenerg

Nuclear

PubHealth
HealthHivAidsProgramPrevent

Reg_LatAmerica
CubaBlockadEconTrade
NucWeapDisarmTreat HealthDiseasPreventControlDevelop CubaNamEconGovRight
CubaEmbargoLawIntern
NicaragGovLatinHondSalvPeople

Table 3.7: Collapsed Topics (11-15)

92

IntPeaceSec
IntCoopOrganPeacSecur
RefugDisplSportPeac
PeacConfSolutSettlementNegot
ConfSecurPreventCoopOSCregion

Noise
DigitLibDagHammarsk NucAgenIAEAenerg

Nuclear

PubHealth
HealthHivAidsProgramPrevent

Reg_LatAmerica
CubaBlockadEconTrade
NucWeapDisarmTreat HealthDiseasPreventControlDevelop CubaNamEconGovRight
CubaEmbargoLawIntern
NicaragGovLatinHondSalvPeople

Table 3.8: Collapsed Topics (16-20)

93

This approach provides a potential opportunity to gain insight into the mechanisms for how
countries use their speeches. While there is no way to specify a priori a speciﬁc category or type of
speech, the post-hoc deﬁned categories may provide additional validation. In this case, prior work
(Mahmood, 2020b) suggests that these speeches can serve an agenda-setting function, resulting in
certain types of countries being more engaged with the UNGA overall. This should suggest that
powerful and democratic countries should, on average, deliver more normative speeches, delivering
language on what they value and what they believe others should value–with language imploring
action, stating what should (or "must") be done, etc. In Table 3.5, I deﬁne a category of speech
as Aspirational language, which is composed of topics most associated with words like "must",
"will", "hope", "can", etc.

The coeﬃcients of the Dirichlet regression are diﬃcult to interpret without additional context,
though they are provided in the appendix. Figure 3.8 provides a heat map for predicted values
across the 20 topics–holding all other variables at their median (or, for factor variables, at the ﬁrst
value in ascending order)–across GDP Per Capita. Mirroring the results from the maximum ALR
values, CINC score is ﬂipped on the Aspirational topic, though the patterns from GDP per capita
and Nuclear capability are consistent with expectations.

The predictions show the strongest trends across Aspirational speech and the category of Devel-
opment, broadly deﬁned: Powerful countries tend to speak much more on aspirational categories,
while weaker countries tend to speak more on Development. Figure 3.9 provides a similar plot
for Democracy. While subtle, this also suggests that Democracies tend to speak slightly more
on development as well, in addition to speaking in more Aspirational language. There may be
further qualitative work to be done to disentangle why speciﬁc issues might be overrepresented
by democracies, including whether their political structures disproportionately incentivize political
posturing on public goods such as economic development and sustainable development.

94

Figure 3.8: Predicted composition across categories, across GDP Per Capita

Figure 3.9: Predicted composition across categories, across Democracy

95

AspirationalDevelopmentDiplomacyGroupGovernanceHumAsstHumRightsIndependenceInstitutionIntCoopIntLawIntPeaceSecNoiseNuclearPubHealthReg_LatAmericaReg_MidEastReg_RusBlocReg_SouthEastAsiaSocialIssuesTerrorCrime4.43116.23356.76937.29397.81418.21728.64639.11069.840210.521512.0446loggdppcFAOvariable0.10.20.3valuevalue0.10.20.3AspirationalDevelopmentDiplomacyGroupGovernanceHumAsstHumRightsIndependenceInstitutionIntCoopIntLawIntPeaceSecNoiseNuclearPubHealthReg_LatAmericaReg_MidEastReg_RusBlocReg_SouthEastAsiaSocialIssuesTerrorCrime−1.9419−0.9706−0.6197−0.3020.14570.39230.6690.97421.3881.9525UDSmedianvariablevalue0.10.20.10.2value3.4 Does issue prioritization correspond to issue salience?

The compositional patterns of speciﬁcity and diﬀuseness in speech are consistent with the
claim that countries tend to speak on issues that are politically relevant, whether this be via a
larger, more diverse politically relevant audience, or whether it be via greater capacity to actually
execute an ambition on a larger set of issues (each salient to some audience). This implies, though,
that the composition of speech should correspond to some external measure of political relevance:
explicitly, the issues a country speaks on should predict that country’s behavior, if the decision to
speak on an issue is driven by the salience of a given issue politically. More generally, the similarity
of issues two countries speak on should predict a similarity of politically relevant issues for both
countries; it should follow, then, that two countries with similar speech compositions should be
more likely to either cooperate or conﬂict with one another, given the overlap of politically salient
issues.

To this end, I use involvement in an ongoing civil war as a measure for whether countries have
overlapping interests, following the framework of Lemke & Regan (2004) and others, who argue
that involvement in a civil war should follow Singer’s (1963) model of inter-nation inﬂuence. From
this, we should expect that involvement in civil wars should–on average–in part be predicted by
whether the potential intervener has similar political interests as the country in conﬂict, making it
more likely that the potential intervener will have a political interest in the outcome (for example,
whether the government is successful or overthrown).26

In order to test this data, I compile a dataset of all civil wars, drawing from UCDP/PRIO’s
Armed Conﬂict dataset, which classiﬁes any organized violence involving the government of a state
as a primary party, and which reaches a 25 battle deaths per year and per conﬂict dyad (Pettersson
et al., 2019; Gleditsch et al., 2002). I include those conﬂicts classiﬁed as either a Civil War or an
internationalized civil war in this dataset. Next, I create a country-conﬂict dataset, allowing every
26This closely mirrors Lemke & Regan’s (2004) hypotheses, which focus on whether countries
are contiguous; whether they have a historical colonial relationship; and whether they are allies
(p.153).

96

country in the system in the year prior to the conﬂict start to be potentially involved in a given
conﬂict. Finally, I code involvement in a civil war as being a binary variable of whether a country is
included as an actor in the civil war by the UCDP/PRIO dataset; I exclude the involved government
as a potential actor in each conﬂict.

Drawing from previous literature, I include several additional variables to model the probability
of involvement. First, I include conﬂict-speciﬁc characteristics such as whether the conﬂict was
classiﬁed as intense by UCDP/PRIO (1000 battle deaths), and the type of incompatibility (gov-
ernment or territory) as classiﬁed by UCDP/PRIO. Second, I capture actor- and location-speciﬁc
factors including the level of democracy (as measured by the Uniﬁed Democracy Scores (Pemstein
et al., 2010)), the total Population, and the GDP Per Capita of both the potentially involved country
and the location country. Finally, I include dyadic measures of whether the two actors (location and
potentially involved country) are joint-members of a United Nations group, the distance between
their respective capitol cities, the colonial history (if any) between the two countries, whether the
two countries are involved in an alliance together, and the diﬀerence in UN General Assembly
Ideal points as measured by Bailey et al. (2015). I also include a log transformed measure of the
maximum number (in any year of the conﬂict) of refugee or refugee-like people originating from
the location in conﬂict to each individual potentially involved country, and a categorical variable
of the world region for the conﬂict location.

To measure the similarity and diﬀerence in which issues countries are speaking about, I use
the ALR-transformation of the yearly-aggregated 81-topic compositional variable, as represented
in Equation 3.5. Given two of these transformed variables (∗
) for a potentially
involved country  and a government experiencing civil war , we should expect that the likelihood
of country –all else being equal–being involved in the conﬂict involving  should increase as
the diﬀerence between the two compositional variables decreases. Substantively, if UN General
Assembly speeches speak to issues that are politically salient to a given country, then as the
composition of issues under discussion becomes more similar, relative to a government
 in
I use the euclidean
conﬂict, an external country  should become more likely to be involved.

 ) and (∗

97

Figure 3.10: Descriptive relationship between Euclidean distance in speech composition and civil
war involvement

distance between the two compositional variables (∗
) as the quantitative measure
for this diﬀerence.27 The full dataset spans across 89 conﬂicts, with covariates coded as their
respective values in the year prior to the starting year.

 ) and (∗

Figure 3.10 provides a descriptive overview of country-conﬂict involvement across all country-
conﬂicts; on average, there emerges a clear pattern that as compositional diﬀerences increase, the
likelihood of involvement in a conﬂict decreases.

I estimate a logistic regression using the variables above; Table 3.9 provides the output for
these models, including a version excluding population, GDP per capita, and Refugee counts in
order to reduce the amount of missing data. The models provide strong support for the claim that
the composition of issues spoken about at the UN General Assembly captures underlying political
incentives that translate into behavior; Figure 3.11 plots the marginal likelihood of a country’s
involvement in a civil war across values of Euclidean diﬀerences. While there are few observations
27Given that some countries only deliver grouped statements in a given year, some of the distances
ended up as zero; the models presented here are ones where I replace zeroes with the minimum
euclidean distance in the dataset. However, all results are substantively identical when we use the
un-corrected distances.

98

0.000.250.500.751.000102030Euc Distance (X axis) vs. Intervention, with smooth (middle) and density (bottom)−0.030.000.030.060.0901020300.0000.0250.0500.0750.1000.1250102030intervention01Figure 3.11: Predicted probability of involvement, across values of Euclidean Distance

at the lowest levels, the changes are dramatic, with the probability of involvement decreasing by
almost 60% as Euclidean distance increases from the 25th percentile to the 75th percentile in the
data.28

3.5 Discussion and conclusion

In the diverse arena of international politics, countries are able to speak on as few or as many
issues as politically beneﬁcial. While prior work has tended to cast skepticism on whether these
types of speeches contain useful information, this paper links the decision to speak on an issue with
a politically relevant audience willing to listen. From this framework, it follows that countries with
a larger, more diverse cohort of politically relevant audiences–such as higher population or more
democratic countries–will tend to speak more diﬀusely, across a larger variety of issues. Moreover,
a country’s capacity to actually act on an issue yields a greater political beneﬁt to agenda setting
on those issues, creating additional incentives to speak towards these interested audiences. This
28Listwise deletion of all observations where Euclidean distance is less than or equal to 7 yields
similar results; this threshold is arbitrary, but determined visually based on the distribution of
distances across the full dataset.

99

(Intercept)

MaxConfIntensity

incompatibilitytype_2

g_ANY_JNT

P5

P5_JNT

Region_LOCAsiaOceania

Region_LOCEurope

Region_LOCMENA

Region_LOCSubSaharAfr

UDSmedianL1

UDSmedianL1_LOC

UDSmedianL1_BothAboveZero

LogTotalRefugeeLike

kmdist

log_FAO_GDPpc_mmL1_LOC

log_FAO_GDPpc_mmL1

log_FAO_tpopL1_LOC

log_FAO_tpopL1

colhist

atopally

UNGAip_DIFF

eucdistCORR_WTwcALRL1_DIFF

AIC
BIC
Log Likelihood
Deviance
Num. obs.
∗∗∗  < 0.001, ∗∗  < 0.01, ∗  < 0.05

Baseline
−1.52
(2.64)
3.09∗∗∗
(0.32)
0.14
(0.31)
0.03
(0.35)
−0.17
(0.48)
1.11
(1.18)
−10.19∗∗∗
(1.20)
−7.82∗∗∗
(1.28)
−2.98∗∗∗
(0.70)
−3.53∗∗∗
(0.64)
0.18
(0.17)
1.07∗∗∗
(0.26)
0.19
(0.32)
0.02
(0.05)
−0.00∗∗∗
(0.00)
−1.96∗∗∗
(0.21)
0.03
(0.10)
0.45∗∗∗
(0.12)
0.35∗∗∗
(0.06)
0.87
(0.55)
1.46∗∗∗
(0.29)
0.23
(0.16)

904.90
1063.71
-430.45
860.90
10085

Baseline_Minimal

−5.51∗∗∗
(0.60)
2.29∗∗∗
(0.22)
−0.01
(0.23)
−0.22
(0.30)
1.05∗
(0.43)
0.48
(1.18)
−2.40∗∗
(0.82)
−4.32∗∗∗
(1.10)
−1.75∗∗
(0.65)
1.46∗∗∗
(0.43)
0.09
(0.14)
0.28
(0.24)
0.79∗∗
(0.29)

−0.00∗∗∗
(0.00)

0.65
(0.54)
1.41∗∗∗
(0.25)
0.05
(0.15)

1072.62
1195.62
-519.31
1038.62
10248

EucDist
2.00
(2.73)
3.20∗∗∗
(0.32)
0.21
(0.31)
−0.36
(0.36)
−0.22
(0.49)
0.83
(1.16)
−10.57∗∗∗
(1.24)
−7.91∗∗∗
(1.29)
−2.46∗∗∗
(0.67)
−3.39∗∗∗
(0.63)
0.12
(0.18)
1.10∗∗∗
(0.26)
0.10
(0.33)
0.03
(0.05)
−0.00∗∗
(0.00)
−2.08∗∗∗
(0.21)
0.07
(0.10)
0.47∗∗∗
(0.13)
0.32∗∗∗
(0.06)
0.92
(0.57)
1.15∗∗∗
(0.30)
0.29
(0.17)
−0.20∗∗∗
(0.04)
880.62
1045.97
-417.31
834.62
9791

EucDist_Minimal

−3.41∗∗∗
(0.80)
2.29∗∗∗
(0.22)
0.05
(0.23)
−0.43
(0.31)
0.96∗
(0.43)
0.27
(1.16)
−2.41∗∗
(0.82)
−4.31∗∗∗
(1.11)
−1.49∗
(0.63)
1.62∗∗∗
(0.43)
0.11
(0.14)
0.21
(0.24)
0.78∗∗
(0.30)

−0.00∗∗∗
(0.00)

0.65
(0.55)
1.16∗∗∗
(0.26)
0.13
(0.15)
−0.15∗∗∗
(0.04)
1055.97
1185.67
-509.99
1019.97
9948

Table 3.9: GLM models: Conﬂict involvement

100

builds in previous work by the author as well, where similar political origins also drove the volume
of speech delivered at the UN General Assembly.

Additionally, the ﬁnding about civil war involvement suggests that in addition to fulﬁlling
political incentives, the liberal argument that political audiences (or constituents, be they domestic
or international) can drive political behavior also ﬁnds support. The composition of speeches is
strongly related to patterns of how countries relate in terms of the diverse issue space in international
politics: those countries with similar politically salient interests will tend to ﬁght the same battles.
In this case, a government experiencing a threat from within will attract outside actors who have
what Singer (1963) refers to as preferred policies surrounding the issues most politically relevant
to that government. Future work can tease out whether patterns of alignment on a given issue is
related to which side an external actor joins in a civil war, even as it is shown here that these actors
will tend to join (regardless of bias for or against the incumbent government) when similar political
issues are discussed in public speeches.

This has implications for the study of diplomacy and international organizations more broadly.
A public forum is one of the most prominent functions that an international organization can
provide. The UN General Assembly is one of the most visible forums in the world; given the
diversity and proliferation of speeches on that forum, the question of whether these speeches are
meaningful becomes relevant to policymakers and scholars seeking to more concretely understand
the interests of world actors. Moreover, it also provides an opportunity to generalize beyond the UN
General Assembly: be it other inter-governmental organizations, or be it worldwide media outlets,
there are a multitude of forums on which world leaders deliver speeches and statements, providing
an opportunity to understand how countries prioritize issues in the context of the possible issues
they can address on these forums.

101

CHAPTER 4

COMMUNICATING FOREIGN POLICY ALIGNMENT VIA STATEMENTS OF

OPINION AT THE UNITED NATIONS GENERAL ASSEMBLY

In September of 2013, Hasan Rouhani gave what the New York Times called "the most widely
awaited speech" at the United Nations General Assembly.1 This speech was widely received as
signaling a potential transition in Iranian foreign policy–particularly toward the United States, its
allies, and Israel–and more speciﬁcally a transition in the direction of moderation.2 This perception
of moderation, notably before Rouhani had taken any tangible policy actions, resulted in a historic
opening of diplomatic between the United States and Iran via a 15 minute phone call between
the two countries3–and led in part to the capacity for Rouhani to negotiate a historic international
agreement on his country’s nuclear program in a way his more hardline predecessor could (or
would) not.4

The co-incidence between Rouhani’s shift in speech with an eventual foreign policy change
points to a deﬁcit in the study of diplomacy in international politics. Namely, it suggests that
these types of speeches–otherwise innocuous statements of opinion, interests, and priorities–can
communicate important information about the landscape of foreign policy interests among nations.
In this case, the shift in Rouhani’s speech aligned him closer with the US and the rest of the
international community, relative to his predecessor whose rhetoric tended to diﬀerentiate him from
them. This seems at odds with a range of previous literature, which has suggested that statements
of this nature by politicians at the United Nations likely do not contain useful information (Voeten,
2005; Farrell & Gibbons, 1989b).

1"Iran’s New President Preaches Tolerance in First U.N. Appearance". New York Times. Rick

Gladstone. 24 September 2013.

and Ed Pilkington. 24 September 2013.

2"Obama tells UN that diplomatic path must be tested with Iran". The Guardian. Julian Borger

3"Obama holds historic phone call with Rouhani and hints at end to sanctions". The Guardian.

Dan Roberts and Julian Borger. 28 September 2013.

4"Former President Ahmadinejad reacts to Iran nuclear deal". Iran Front Page News Service.

16 July 2015.

102

The question of how speech patterns could potentially map to future behavior is of critical
importance. Especially on high-stakes issues such as nuclear proliferation, policymakers and
diplomats would beneﬁt from understanding how to interpret statements of opinion, which are
otherwise proliﬁc in international politics and often occur on a continuous basis. For scholars
of international relations, understanding this type of information in a systematic way allows for a
more expansive study of diplomacy and international communication–particularly at international
institutions such as the UN.

In this paper, I examine the conditions under which speeches at the UN General Assembly can
communicate useful information about the alignment pattern of foreign policy interests. I argue
that leaders face a tradeoﬀ between the the cost of aligning with unfriendly peers and the incentive
they have to pander to multiple audiences. An analysis of UN General Assembly speeches between
1984 and 2012 shows that on more controversial issues, where a history of disagreement separates
between friends and foes more easily, the alignment of speech patterns map most clearly on to
the alignment of behavioral patterns, suggesting that in these cases leaders restrict themselves to
speaking mostly with friendly peers.

4.1 Speech, foreign policy alignment, and the United Nations

A rich literature exists in international relations on how states communicate. From the cred-
ibility of threats (Schelling, 1960; Waltz, 1979; Guisinger & Smith, 2002; Fearon, 1994, 1995;
Sartori, 2005; Kydd, 2005; Press, 2005), to the alignment of policy interests (Morrow, 1994, 2000;
Crescenzi, 2007), to the salience of a demand (Trager, 2011), there are a range of things that leaders
need to communicate to other state leaders. This communication generally takes the form of either
actions or speech, such as conﬂict and troop deployment in the ﬁrst category, or verbal threats and
verbal demands in the second category.

While the scholarly interest in these types of communication has grown over the past several
decades, a speciﬁc form of communication has remained largely under-theorized: public expres-
sions of opinion and sentiment. This is particularly important since these types of speeches are so

103

proliﬁc in politics writ large: from media interviews to the United Nations, state leaders communi-
cate an almost continuous stream of opinions and sentiment over time to a host of audiences on a
range of issues. While there has been a fair amount of work done on communication within states
by political elites (Monroe et al., 2008; Quinn et al., 2010; Proksch & Slapin, 2012), including
some theorizing on the data generating process thereof (Budge et al., 1987), there has not been
an analogous eﬀort in international politics to understand speeches between states. Not only is
this type of speech proliﬁc in international politics, its informative value is often assumed in many
scholastic studies of international politics, either in case study narratives (Press, 2005; Thompson,
2009) or as a variable of interest which aﬀects other processes Thyne (2009); Hayes & Guardino
(2013).5

This informative value of speech, however, has not been subject to systematic testing. More
explicitly: what, if any, information is being communicated in these expressions of opinion and
sentiment? To better understand this, we can look to the literature on domestic party politics.
Proksch & Slapin (2012), for example, argue that public debates serve to "communicate...policy
positions to their parties, other parties, and voters", arguing that parties have incentives to "maintain
[their] brand" by "preventing [members] from undertaking activities that contradict the party’s
primary message" (p.521-522). In this way, the choice to prioritize some issues over others – for
instance, emphasizing symbolic and social issues near an election (Quinn et al., 2010) – or the
choice of one "brand" over another "brand" (in terms of the sentiment they express on an issue) can
serve an informative purpose.6

5Ramsay (2011); Bils & Spaniel (2017) are notable exceptions; they provide promising frame-
works to understand how uncertainty over the alignment of interests can lead to credible commu-
nication, both in terms of policy coordination and preventing militarized conﬂict.

6This is also similar to Trager’s (2011) theory of multi-dimensional diplomacy, wherein he
argues that during bargaining, leaders can communicate issue salience by prioritizing some issues
over others–for example, choosing to demand a more diﬃcult concession over an easier one,
suggesting that the leader is willing to risk negotiations falling apart and potential war. While
his theory exists in the context of private negotiations and the communication of resolve, it is
directly applicable to public deliberation and the communication of alignment, in that the choice
to emphasize some issues over others reﬂects a prioritization that diﬀerentiates one leader/country
from another.

104

Yet international relations is markedly diﬀerent than domestic politics. Most importantly,
formal political parties do not exist the way they do in countries, and so the eponymous "brand"
analogy made by Proksch & Slapin (2012) breaks down since these brands ultimately do not exist
in a structured way.7 This speaks to the problem that Voeten (2005) correctly identiﬁes with regard
to deliberation at forums like the United Nations, where there are only "a shallow set of common
values... [and] obvious incentives to misrepresent their positions, as the stakes are clear and the
relevant actors few". This would seem to render it diﬃcult to map particular opinions and argument
to some kind of identiﬁable position (p.537).8 Ultimately, without some kind of supra-national
lexicon by which to identify a particular country’s branding, or without a mechanism linking
statements to a clear incentive structure, the conventional wisdom among international relations
scholars is generally that these statements ultimately have little informative value.

However, there are two areas for improvement. First, especially in public arenas like the United
Nations, the number of "relevant actors" are in fact quite large. These pubic arenas broadcast
opinions not only to other state leaders, but also to interested political elites across the world. Be
they NGOs, domestic elites, or elites in other countries, the scope of relevant actors is large and
heterogenous. Since a given speech is ﬁnite in both size and content, a leader must pick and choose
between audiences – creating a strategic environment and incentive structure. Second, while there
is no supra-national lexicon to objectively identify any given speech, there is an endogenously
deﬁned lexicon by which to compare speeches to one another. More explicitly, while the absolute
content of a speech may not be meaningful, the relative content can be: an emergent lexicon exists
as leaders speak on any given issue over time, providing a medium in which to compare any set of
speeches.9

7Lake (2009) provides a compelling counterargument to this, where he argues states do in fact
exist in "hierarchies". An interesting theoretical link exists here in bridging the literature on party
politics and international relations, with respect to messaging and branding.

8For arguments defending the UN’s discursive and persuasive roles, see Johnstone (2003, 2005);

Krebs & Jackson (2007)

9While a more detailed, descriptive lexical analysis of the UN is beyond the scope of this paper,
there is useful work on how the language of politics changes over time (Greene et al., 2019). For a
simple example, consider the Topic Model which I describe later in the paper. One of the uncovered

105

The UN General Assembly is the most natural starting point for analyzing how statements
of opinion and sentiment on foreign policy communicates foreign policy interests, and how this
maps onto observable behavior. Since its inception in 1945, scholars of international politics have
studied the UN from a variety of angles, including its role in legitimizing foreign policy (Claude,
1966; Voeten, 2005; Thompson, 2009; Chapman, 2011), the Security Council’s authorization of
collective foreign policies (Fortna, 2008; Beardsley & Schmidt, 2012), and with respect to UN
General Assembly voting Alker (1964); Alker & Russett (1965); Vincent (1971, 1972); Kim &
Russett (1996); Voeten (2000); Bailey et al. (2015).

The General Assembly in particular is useful here for several reasons. First, it includes all
countries in the world. This membership scope provides a unique opportunity to observe statements
of opinion in a context where all countries speak on similar issues, and in a similar environment with
a commonly understood emergent lexicon at any given time. This makes cross-country comparison
more robust. Second, the wide membership combined with the General Assembly’s high proﬁle
provides a large set of "relevant actors" (to use Voeten’s (2005) terminology) from which leaders
must pick and choose an audience. This creates a strategic environment and incentive structure
whereby leaders must prioritize some interests of some audiences over others. Third, the UN
General Assembly provides a very useful empirical opportunity to map speech patterns to relevant
behavior, issue by issue. In this case, as I discuss in more detail below, the UNGA provides an
opportunity to map speech patterns onto foreign policy behavior across knowable issues by using
voting patterns to measure foreign policy behavior. Finally, the General Assembly is useful from a
theoretical perspective in this study, since it has little if any true legal power. In other words, the
null hypothesis and conventional wisdom – namely, that speech patterns reﬂect a babbling process,
topics is consistently composed of words like "international", "tribunal", and "law", which based
on post-hoc domain knowledge likely maps on to an international law topic. Starting in the middle
of the dataset, toward the early 2000s, the terms "rome" and "icc" begin to appear. This is almost
certainly due to the fact that the Rome Statute establishing the ICC as an arbiter and player in
international law was adopted (1998) and implemented by 2002. In this way, the emergent lexicon
surrounding the issue of international law shifted slightly to reﬂect changes in the relevant players.
To cite the ICC during an ongoing crisis, then, becomes associated with the politics which surround
the institution–and the countries who invoke it during that crisis.

106

Figure 4.1: Three countries speaking in similar ideological spaces

Figure 4.2: Three countries speaking in very diﬀerent ideological spaces

conveying little or no systematic information about foreign policy interests – should be favored
most strongly in a forum like the UN General Assembly, which might otherwise be seen as merely
a "contemptible talk-shop" (Claude, 1966, p.372).

4.2 A scale of two interests: Alignment through relative content

When leaders speak, audiences listen. Though this axiom has mostly been used to understand
how domestic audiences can make threats more credible (Fearon, 1994; Weeks, 2008; Levendusky
& Horowitz, 2012; Tarar & Leventoğlu, 2013; Kertzer & Brutger, 2016), there is also work to
suggest that it applies more broadly to international audiences as well–both overseas intra-state
audiences (Hayes & Guardino, 2013) and other countries (Sartori, 2005; Ramsay, 2011). In a high-
proﬁle setting like the UN General Assembly, this means that all public actions–such as voting and
speeches–are subject to this simple but important axiom.

What, if any information can these audiences glean from speech, and why would leaders care?
A critical issue for onlookers to resolve from these speeches is that the underlying scale is, a priori,
undeﬁned. For instance, consider three countries, , , and . This issue can be illustrated by
a hypothetical situation where  states that he is a "7" on the issue at hand. Without additional
information, the content of that speech is essentially indeterminate, since as Voeten (2005) points
out there is only a shallow "common set of values" in which to interpret this statement. That is, a "7"
could mean vastly diﬀerent things depending on the context: compare Figure 4.1 and Figure 4.2,
where the same statement maps  on to very diﬀerent sides of the ideological scale (holding  and
 constant in their own positions).

107

Drawing from the terminology used earlier, this problem occurs from reading the absolute
content of speeches: a "7" has little to no meaning in an absolute sense without some clarifying
set of values in which to interpret it. However, as Figure 4.1 and Figure 4.2 allude to, ’s speech
does still contain relative content. That is, it remains possible to make robust statements about
Ɗ, Ɗ, and Ɗ. For a more concrete example, if  and  emphasize their "concern about
the humanitarian situation" in some civil war , and  tends to emphasize his concern about
"maintaining institutional stability" in the same civil war , while it is diﬃcult to say whether , ,
and  are diﬀerent in some objectively meaningful way, it is much easier to say that Ɗ is smaller
than Ɗ–meaning that  and  are more aligned in their foreign policy speech. The informative
nature of diﬀerence and similarity is not a novel ideal in the study of diplomacy: Jervis (1970,
p.21) argues a similar idea, noting that "all secret messages and most public ones are signals...these
signals may involve following an established routine, as in the use of diplomatic language, or
may entail breaking that pattern" (emphasis added). This pattern is deﬁned by the community
at large, or by a sustained pattern over time, subject to change.

Interestingly, under this framework the information about alignment in speech depends only
on content, and requires no additional statements about the truth or falsehood of that content. For
instance, whether or not  is "truly" concerned about the humanitarian situation in  has no bearing
on the fact that in his speech,  is communicating a closer alignment to  than he is to . This is
a critical takeaway from the theory presented here: there might be valuable information contained
in speeches even if leaders are not faithfully representing their true interests. The emphasis on
"true" here is important, since it is often a red herring that can obfuscate the process of gleaning
information when leaders deliver public statements of opinion.10

Also noteworthy is that the scales are deﬁned at the bounds by the most polarized content:
10In the extreme example, suppose  is never communicating their "true" opinion on a given
issue. The logic of alignment presented here would still apply, since the content of a speech can
still be assumed to represent something that a corresponding audience wants to hear. For instance,
a verbal threat may not always result in a conﬂict, and may not always represent the resolve or
capacity to ﬁght, but it would be assumed that consistently delivering verbal threats still reﬂect a
level of hostility above the absence of any verbal threats (Terechshenko, 2020).

108

given some endogenously deﬁned vocabulary, countries are identiﬁed in alignment relative to one
another in a space deﬁned by those countries who speak most diﬀerently than one another on that
particular issue. This leads to interesting theoretical opportunities to study patterns of polarization
more broadly in international politics, though this is beyond the scope of this paper.

4.2.1 Mapping speech to friends and foes

Political actors care about other leaders’ interests. For domestic actors, there is a lingering concern
that leaders will forego the interests of the public in order for a leader to pursue his/her own private
interests (Stasavage, 2004; Svolik, 2012; Colaresi, 2014), or more generally that a leader may
forego the interests of one domestic group in favor of another. For international actors, a similar
concern applies: decisions about foreign policy making are driven by the anticipation of what
other leaders will do (Findley & Teo, 2006), and more speciﬁcally by the anticipation of whether
these other leaders will implement favorable (or "preferred") policies (Singer, 1963). Learning
about the alignment of interests becomes a vital part of coordinating foreign policy, particularly
when the stakes involved with pursuing a diﬃcult or risky policy–where counter-action might be
extremely costly (Gent, 2007)–are very high. As Bils & Spaniel (2017) argue, uncertainty over
ideological positions can make cheap talk a useful component of pre-crisis bargaining and the
ultimate decision to go to war. They show that the presence of many potential opponents with
disparate policy preferences can render cheap talk eﬀective, and more importantly that the variance
of perceived alignment in latent ideological space can have critical implications–including, in
some cases, whether war breaks out. Thus, there is wide agreement in the scholarly literature that
the audiences who are listening to this relative content care about (and can potentially use) the
information they glean.11.

Connecting relative content to alignment, this creates a decision for leaders to make. Consider,
for example, a simple setting where two interested onlookers– and –are watching  and .
11Also see (Ramsay, 2011), who argues that pre-crisis signals about potential alignment of

interests can aﬀect peace by leading to bargaining in the ﬁrst place

109

Assume that  and  disagree on some issue , and that their behavior reﬂects a decision to reﬂect
the interests of  or  respectively, whose preferences are conﬂicting on issue . There are
two possibilities with regards to the speech patterns of  and  on issue , given their conﬂicting
behavior on that issue. The ﬁrst possibility is that  can speak in a way that also diﬀerentiates him
from the way  speaks on that issue. Under this scenario,  is using his speech to double down, or
reinforce his disagreement with . The second possibility is that  can speak in a way that aligns
him with the way  speaks on the issue. Under this scenario,  is using his speech to hedge against
his disagreement, or pander to the interested onlooker  who otherwise sees  siding with a foe
in their behavior.12

4.2.1.1 Empirical implications

Given this, we should expect that patterns of reinforcement and hedging are driven by two things.
First, as the distance increases between some country ’s observed behavior (for example, a weak
country  being coerced into supporting the interests of ), and the interests of some politically
important constituent  that  values, we should expect that these leaders will beneﬁt more from
hedging at all–since they will have a constituent to whom they must pander verbally given a lack
of alignment in their behavior. For example, consider the case of voting on UN General Assembly
resolutions, where a range of prior work has elucidated the role of vote buying–especially by more
powerful countries against weaker countries (Carter & Stone, 2015; Dreher et al., 2008; Wang,
1999; Rai, 1980). Here, if a given country  has some underlying foreign policy interest , the
weakest countries would be most vulnerable to being pulled away from that position and towards
12While this can be expressed in terms of dyadic relationships between countries, there is
no reason we cannot equally express this in terms of issues themselves. For instance, consider
some issue , with two sides of the issue preferred by some groups  and .
If  and 
in their behavior act diﬀerently than one another, they can use their speech to either reinforce
this diﬀerence (thus speaking diﬀerently than one another, mirroring their disagreement) or hedge
against this disagreement (where either  or  panders to the other side in speech, while acting in
behavior in favor of their side of the issue). The dyadic formulation facilitates observation, but there
may be some implications to the issue-focused interpretation–for instance, whether the complexity
of an issue causes diﬀerences in how countries align with one another.

110

the position expressed by some resolution or action. That is, if a resolution’s position is expressed
as , then the magnitude of  −  should be much higher for the weakest countries than the
strongest countries. As a result, we would expect that the weakest countries will be most likely to
hedge their behavior.

Hypothesis 1: As the power diﬀerence between two countries increasing, their speech
will be less reinforcing

Second, we should expect the tendency for reinforcing or hedging speech to change depending
on the political conditions between two countries. Speciﬁcally, given a pre-existing relationship
between countries, the capacity to separate between friends and foes should become more clear–
resulting in an increased cost of aligning with one or the other inconsistently. For example, if
two countries with a hostile rivalry vote diﬀerently on an issue at the UN General Assembly, the
cost of hedging against that vote for either country will be higher than if the rivalry did not exist.13
Similarly, if two countries have political constituents who tend to favor one another, hedging against
an agreeing vote could be costly.14 In this case, we can use Rivalry to measure pre-existing enmity
between countries, using data from Thompson & Dreyer (2011). Second, I use joint Democracy
to measure pre-existing camaraderie, following Tomz & Weeks (2013), who show that politically
13In some ways, this mirrors recent ﬁndings in the literature on Audience costs. As Kertzer
& Brutger (2016) argue, constituents will punish leaders for being inconsistent with their (the
constituent’s) political propensities: hawkish constituents with pre-existing beliefs about the use of
force will punish leaders for inconsistency, while dovish constituents with pre-existing beliefs about
paciﬁsm will punish users for the act of threatening in the ﬁrst place. Given that rivalries have been
shown to have deep roots in domestic politics (Colaresi, 2005), in the rare cases when rivals agree
on an issue, hedging against that agreement would risk raising questions about why leaders still
chose to express agreement through behavior; similarly, when rivals disagree on an issue, the cost
of hedging against that disagreement risks raising questions about whether the leader is actually
disagreeing to a suﬃcient degree with a known rival. Hence, we should expect less hedging. There
may be future implications here for studying the dynamics of rivalries, since the prediction of this
theory would be that as rivalries de-escalate, hedging would be more of a possibility since leaders
will be balancing conﬂicting incentives to reconcile with a rival while also still maintaining the
rivalry as politically necessary.

14Much like with rivalries, we should expect that given a known preference between two polit-
ically relevant groups, the cost of hedging against behavior should increase since friends and foes
are more clearly deﬁned

111

relevant Democratic constituents tend to view other Democracies as sharing their interests, more
worthy of trust, and more worthy of moral consideration with respect to war.15

Hypothesis 2: Rival countries will reinforce in their speech more

Hypothesis 3: Joint Democracies will reinforce in their speech more

Third and ﬁnally, we should expect that the conditions surrounding an issue at the UNGA can
shape the incentives surrounding alignment patterns of both speech and behavior. Speciﬁcally, on
the most controversial issues, there will be a consistent pattern of disagreement among nations:
on these issues, the prevalence of disagreement creates the ability to separate between friends and
foes. For example, on the issue of the Middle East, there is a consistent pattern of disagreement
between countries on multiple possible questions. This creates separation: determining whether
Libya and Iran are a friend or foe to the United States on this issue is relatively simple. Similarly, re-
turning to Hassan Rouhani’s speech in 2013, his rhetoric on the nuclear issue represented a marked
change in alignment patterns from his predecessor, calling international concerns "reasonable"
and stating "unambiguously" that Iranians sought peaceful technology and embraced "international
cooperation" and "transparency".16 Compare this to rhetoric in 2012, where on the nuclear issue
then-president Mahmoud Ahmadinejad focused on "intimidation by nuclear weapons...by the hege-
monic powers", the "continued threat by [Israel]", and the "state of mistrust...[where] no one feels
safe."17 This shift was radical in its stark diﬀerentiation from his predecessor and its move toward
reconciliatory language, communicating information on the new regime’s priorities and foreshad-
owing the eventual cooperation that ended up occurring.18 More importantly, the high-stakes nature
of this issue meant that the combination of a radical shift in rhetoric and the intense polarization

15Tomz & Weeks (2013, p.857; p.860)
16Transcript accessed from The Times of Israel website. Accessed December 2018.
17Full transcript available here.
18Interestingly, the subsequent termination of the Iran agreement by the Trump administration
in May of 2018 (source: BBC online) is also predictable from another shift in rhetoric, though this
time originating from the United States.

112

that had arisen around the issue created an environment where this shift in language communicated
a substantial amount of information to the international community.

Thus, on these controversial issues, there is much more possible information to be gleaned from
speeches, in that when  speaks then  and  are more readily able to determine whether 
is speaking similarly or diﬀerently than friends and foes on that issue. Contrast this with much
less controversial issues, where the information is more muddled: when nations generally agree on
an issue–such as world peace in more abstract terms–the diﬀerences in Ɗ or in Ɗ are less
informative, since it remains unclear who is a friend or foe of whom (or if there are any foes at
all). Further, in these cases politicians should have a less restrictive environment in which to speak
freely–for instance, to pander to multiple audiences, or to explore the ideological space more freely.
This should result in more randomness with regard to who countries are aligned with at any point in
time, regardless of their eventual behavior. Thus, controversial issues should be more informative
overall, in that speeches in these environments should map most closely with eventual behavior.

Hypothesis 4: Higher controversy issues will see the most reinforcing speech with
respect to behavior.

Statistically, we can represent these tendencies of reinforcement or hedging as the relation-
ship between the alignment of speech patterns with the alignment of behavior. More speciﬁcally,
given two countries  and , the most reinforcing speech occurs when as the diﬀerence in speech
Ɗ ( , ) increases, the diﬀerence in their behavioral outcomes Ɗ ( ,  ) also sees a corre-
sponding increase. The most hedging speech is the inverse, where an increase in one corresponds
to a decrease in the other. Thus, given some model:

Ɗ ( ,  ) =  × Ɗ ( , )

describing a linear relationship between Ɗ ( ,  ) and Ɗ ( , ), we can say that as 
increases in value, speeches are becoming more reinforcing with respect to behavioral outcomes.
In this paper, I use voting on UN General Assembly resolutions as a measure of behavior; as

113

discussed in the next sections, this allows for us to model variation across issues–in this case,
diﬀerentiating more controversial issues from less controversial issues.

4.3 Empirical strategy

To measure speech patterns, the data comes directly from the United Nations General Assembly,
where I have scraped the transcripts for all plenary sessions in PDF format, translated them to text,
and identiﬁed each speech with its corresponding country. This corpus was collected independently
of the one by Baturo et al. (2017), whose corpus focuses only on speeches delivered by heads of
state under General Debate.19 To measure behavior corresponding with those speeches, I rely on
the raw text of UN General Assembly resolutions; I use a scraper built by Erik Voeten20 to extract
the full text of resolutions starting in UN General Assembly session 48, corresponding to the year
1993, and I collect the remaining resolution text back to 1984 myself. Thus, the time frame for this
paper is 1984-2012.21

4.3.1 Mapping speeches to resolutions

As discussed in a later section, one of the variables of interest in this paper is the controversy
surrounding an issue. This requires separating both speeches and resolutions into categories
of issues, and classifying each into corresponding groups. Mapping speeches to corresponding
resolutions is a complicated task. For example, a resolution can be on diﬀerent topics, from
nuclear proliferation to the Middle East to a UN humanitarian mission. However, a single speech
can encompass multiple topics. This makes it diﬃcult to generalize from a broad, multi-topic
speech to a single resolution. To rectify this problem, I separate each General Assembly speech
into paragraphs. Next, I use a Structural Topic Model (STM) (Roberts et al., 2014) to classify
19For example, their data from the year 2000 (session 55) includes the speech made by Madeleine
Albright in the 10th plenary meeting of the 55th session, but leaves out President Clinton’s address
during the 3rd plenary meeting, which occurs outside the conﬁnes of "General Debate" and under
diﬀerent agenda topics.

20Email correspondence, 5 May 2017
21Voting data is only mapped to resolutions through 1984; see previous email correspondence.

114

paragraphs and resolutions into distinct topics. STMs classify documents – in this case, paragraphs
and resolutions – into topics via a mixed-membership model, where a document can be composed of
many topics at once. The model does so by estimating ( |) across all words, allowing
for post-processing by the analyst to label topics (Quinn et al., 2010).

Choosing the number of topics is a diﬃcult task, with no correct answer (Grimmer & Stewart,
2013). To rectify this, I marry together the quantitative and qualitative methods by estimating a
high-dimensional topic model–with the speciﬁc number of topics decided upon by the algorithm
proposed by Mimno & Lee (2014)–and then grouping these topics into broader categories by hand.
The STM model returned 87 topics, which I then further grouped based on domain knowledge of
UN General Assembly speech content, as well as from experience running topic models of lower
dimensionality.22 For example, given two topics {Millen, Develop, Goal, Sustain, Commit} and
{Econ, Povert, Environ, Soc, Polit, Erad}, I collapsed both into a single category of Development.
Similarly, I collapsed seven (7) topics containing normatively active terms such as "can", "will",
"hope", "aspir", "must", "believe", etc. into a single category of Aspirational.

Through this process, I identify 31 categories that compose resolutions and speeches. To
convert the original 81-topic composition to the 31-category composition, I take the sum of the
proportions returned for all topics which are part of a given category. For example, given the 81
topic composition of a single speech or resolution, in order to ﬁnd the proportion on the category of
Development, I take the sum of the proportions given in the topics {Millen, Develop, Goal, Sustain,
Commit} and {Econ, Povert, Environ, Soc, Polit, Erad}. This results in each speech and resolution
being represented by a vector of length 31, which sums to 1.00. Finally, I classify each resolution
and paragraph into the category with the highest proportion.23

22For instance, whether 20, 30, or 40 topics, topics such as Nuclear Non-Proliferation, Aspi-
rational goals, Human Rights, and Economic Development were always present; these are also
examples of issues which a subject matter expert on UNGA speeches would expect to ﬁnd in the
underlying content.

23For example, a 20-topic model might return a vector [0.0001, 0.0001, 0.0001, 0.0001, 0.0001,
0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.11, 0.0001, 0.0001,
0.0001, 0.23, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.65,
0.0001, 0.0001], representing a single point on the 31-dimensional simplex. In this paper, I would

115

Proportion

0.018
0.004
0.059
0.005
0.097
0.003
0.002
0.002
0.004

0

0.096

0
0.01
0
0.15
0.001

0

0

0.001
0.005

0.001
0.009
0.049
0.04
0.027
0.208
0.002
0.06
0.141
0.005

0

Category
Aspirational

Conﬂict

Development
Diplomacy
Economic
Governance
Humanitarian
HumanRights

Illicit

Independence

IntCoop
IntJustice
IntLaw
MidEast
Misc

Sustainability

UN_GenAssembly
UN_Operations
UN_Principles
UN_Reports
UN_Resolution

UNSC
Violence

NatDisaster

NonProliferation
NuclearSafety

Peace

PoliticalRights
PublicHealth

Social
Speech

Table 4.1: 31-category composition for UNGA resolution A/RES/47/214

Finally, especially given the nature of UN General Assembly resolutions, there are administrative
and institutional categories that are not as relevant here, since they do not reﬂect any major category
of foreign policy interest. For example, the resolution A/RES/47/214 is 13 pages long, and is an
extensive resolution on the proposed medium-term plan for the years 1992-1997 with respect
to programme performance, budgeting, and administrative committees for the UNGA. The 31-
category mixture for this resolution is:

In this case, the three highest proportions belong to administrative categories. For this dataset,
I remove the categories Misc, Speech, UN General Assembly, UN Operations, UN Reports, UNSC.
classify this document into topic 29. I remove any resolutions which are less than 10% composed
of any single topic.

116

I retain the remaining categories as listed in Table 4.1.

4.3.1.1 Measuring alignment of speech

A key independent variable of interest is the similarity between any two countries’ speeches. There
has been a growing literature in measuring similarities and diﬀerences from political text (Grimmer
& Stewart, 2013), including with applications to measuring similarities and diﬀerences between
groups such as countries, political parties, or party manifestos (Budge et al., 1987; Laver et al.,
2003; Slapin & Proksch, 2008; Monroe et al., 2008; Elﬀ, 2013). To measure diﬀerences between
countries on speech, I draw upon previous work, drawing from the aforementioned literature and
treating words as data, to estimate ideal point estimates drawing from the text of the UN speeches
(Mahmood & Colaresi, 2016, 2017). I estimate ideal point measurements using a dynamic linear
model in the form:

 
(×1)

 
( ×1)

= zt

 
(×)
( ×1)
= A( ×) −1
( ×1)

+  
(×1)
+  
( ×1)

(4.1)

where  represents each word spoken in the full vocabulary of all speeches and  represents
the number of dimensions in which the ideal point estimates  are estimated per country over time.
In this model,   represents a noisy projection of latent ideal points   in -dimensional space,
where the state equation represents the evolution over time of  , for country  at time . I use
Euclidean distance as the distance metric between the estimated positions; in the main body of the
paper, I present results from a 2-dimensional model.24

24I also run all analyses excluding the United States and Israel from the analysis, given their
unique role with the topic of Israel at the UN over the past 70 years. I also run these models using
both 2- and 3- dimensional ideal point models. I also estimate the results using Proksch & Slapin’s
(2012) WordFish model instead of the one presented above. Since their model does not include a
dynamic component, I estimate a diﬀerent model per year across all countries. All results remain
substantively similar.

117

4.3.1.2 Measuring levels of controversy

Recall that one of the predictions of the theory is that the direction of relationship between
speech patterns and voting patterns depends on the strategic incentives present on a given issue. To
operationalize these incentives, I use the proportion of dyads who disagreed at all on any resolution,
constructed for each topic, to rank order topics every year by their respective levels of controversy
(using voting patterns from the previous year).25 This provides a scale of least to most controversial
issues.

To justify this, consider the role of controversy in the theory presented earlier. Controversy
speciﬁcally pertains to the ability to identify two countries  and  as either tending to be friends
or tending to be enemies. This is a diﬃcult concept to measure in any objective way. While a
thorough discussion of identifying objective enmity or friendship between countries is beyond the
scope of this paper, I focus primarily on the UN as a way to most closely approximate friends
and enemies in the context of the UN. To do this, I consider each resolution  as representing
a single possible question in the broader array of possible questions that any given issue  could
potentially produce. It is reasonable, then, that as a given issue  sees a larger and larger tendency
for countries to disagree on more and more questions, it can be considered more "controversial".
On these issues, countries are displaying a tendency to have more polarized views–such as the
Palestine issue above–whereas on less controversial issues the sorting between friends and foes is
much less clear. Most relevant to the research design, it also provides a rigorous way to measure a
yearly, issue-by-issue level of controversy at the UN General Assembly over time, mapping those
to corresponding speeches and resolutions in order to test the theory presented earlier.

Figure 4.3 shows the count of resolutions across topics. Figure 4.4 shows the 25th, 50th,
and 75th percentiles of calculated speech alignment on each topic, aggregated across the years.
Finally, using the measure of controversy as described here, Figure 4.5 shows the average level
of controversy–across all years–per topic. Notice that a large number of topics have controversy
25For example, in the year 1995, the measure of controversy would be created by calculating the
proportion of dyads who disagreed, based on all dyads in the dataset within 1994, for each topic.

118

Figure 4.3: Number of resolutions, by topic

Figure 4.4: Speech alignment, by topic

levels of zero: this indicates that there was no disagreement present on that issue. To this end, I run
all empirical analyses including these zeroes, but for robustness I also run models dropping those
issues with zero controversy.

4.3.1.3 Behavioral similarity

Finally, recall that the goal of this paper is to measure whether or not similarities and diﬀerences in
speech map onto similarities and diﬀerences in observed behavior, and the direction in which this

119

Topic_AspirationalTopic_ConflictTopic_DevelopmentTopic_DiplomacyTopic_EconomicTopic_GovernanceTopic_HumanitarianTopic_HumanRightsTopic_IllicitTopic_IntCoopTopic_IntJusticeTopic_IntLawTopic_MidEastTopic_NatDisasterTopic_NonProliferationTopic_NuclearSafetyTopic_PeaceTopic_PoliticalRightsTopic_PublicHealthTopic_SocialTopic_SustainabilityTopic_Violence0200400600Resolution count (all years)TopicTopic_AspirationalTopic_ConflictTopic_DevelopmentTopic_DiplomacyTopic_EconomicTopic_GovernanceTopic_HumanitarianTopic_HumanRightsTopic_IllicitTopic_IntCoopTopic_IntJusticeTopic_IntLawTopic_MidEastTopic_NatDisasterTopic_NonProliferationTopic_NuclearSafetyTopic_PeaceTopic_PoliticalRightsTopic_PublicHealthTopic_SocialTopic_SustainabilityTopic_Violence0510Dyadic Speech Distance (Median and IQR)TopicFigure 4.5: Controversy level, by topic

mapping occurs (inverse or direct). The dependent variable, therefore, is the similarity of voting
on some resolution between any two countries  and . This requires some speciﬁc decisions,
however. There are 4 possible ways a country can act on a particular resolution: a yes or no vote,
an abstention, or an absence. To collapse this into a binary decision, I code abstentions as no votes,
and I treat absences as missing data.26 Finally, since the purpose here is to predict similarity and
diﬀerences in votes, I convert these to an [undirected] dyadic dataset covering all dyads, where {1}
indicates disagreement in voting within the dyad and {0} indicates agreement in voting. Given that
the covariates of interest (controversy, speech alignment) vary at the level of topic-year, I collapse
all observations to the topic-year unit of analysis, with the dependent variable of interest being the
count of disagreements and agreements on all resolutions within a particular dyad.

4.3.2 Research design

The research question is concerned with the conditions under which countries reinforce or hedge
in their speech, and more speciﬁcally the pattern by which alignment in speech maps on to an
alignment of interests. Rather than being a causal question, then, this is a measurement question:
26The raw voting data also includes a placeholder for countries that are not members of the

international system at the time of voting. I treat these as missing as well.

120

Topic_AspirationalTopic_ConflictTopic_DevelopmentTopic_DiplomacyTopic_EconomicTopic_GovernanceTopic_HumanitarianTopic_HumanRightsTopic_IllicitTopic_IntCoopTopic_IntJusticeTopic_IntLawTopic_MidEastTopic_NatDisasterTopic_NonProliferationTopic_NuclearSafetyTopic_PeaceTopic_PoliticalRightsTopic_PublicHealthTopic_SocialTopic_SustainabilityTopic_Violence0.000.050.100.150.20ControversyTopicgiven a causal relationship between interest patterns and behavioral patterns, can the alignment
of speech patterns provide empirical leverage in explaining the alignment of these underlying
interests, and what are the conditions under which they can do so?27. Given the hypotheses above,
such conditions should include the level of controversy of a given issue in a given year; the presence
of a rivalry between two countries, and whether or not the Cold War is ongoing. To model this, I
use a multiplicative interaction:

  =   

  +   

 
+ 

  ∗  
 

(4.2)

(cid:16)

(cid:17) +   +  

The interpretation of coeﬃcients in this case should not be treated as a causal one, but rather
as a statistical one: namely, whether and how variation in patterns of behavior are explained by
variation in patterns of speech.28 More speciﬁcally, reinforcing speech here will be represented as
a positive relationship between  
–and, more importantly, that
this relationship is speciﬁcally more positive under some conditions versus others. For example,
given some binary variable of whether a given pair of countries are in a rivalry, when  
= 0
to represent no Rivalry, the relationship between  
  and   (that is, the alignment of speech and
the alignment of behavior) will be given by , and we should expect this to be less than than the
corresponding Rivalry pairs, when  

  and  , at some level of  

= 1, given as  + .

 
27Acute observers will note this is a somewhat ﬂipped version of instrumental variable estimation,
where one is given a useful measurement of an underlying concept in order to evaluate the underlying
causal relationship

28This is a crucial distinction, especially given that speeches on speciﬁc issues can occur both
before and after a vote on that issue. Countries can raise an issue in speech after a vote is made
on that issue, and can vote on the same issue multiple times in a single year. The current level
of granularity in this paper is at the issue level of analysis, rather than the resolution level: as a
result, we cannot say for certain that speciﬁc patterns of alignment on a given resolution will predict
agreement or disagreement. Rather, we can only say that there are patterns of correlation between
voting alignment and speech alignment on a given issue, in a given year, and that these patterns
follow expectations derived from theory. Moreover, this paper does not address the question of
persuasion: there are no claims presented here that speech prior to a vote can cause changes in
voting afterward. Thus, the temporal order of speeches are not in question for the purposes of
this paper. Future research can attempt to more concretely identify single speeches with single
resolutions, and tease out both causal and correlative relationships accordingly.

121

In this paper,   represents the disagreement between two countries on UN General Assembly
resolutions. To model this, I assume that   is distributed   ∼ Binomial (, ) for each dyad-year-
topic unit of analysis, and use a logit link function to transform the outcome for linear predictors.

Model speciﬁcation and UN voting blocs

I include several additional indicators for modeling dyadic voting patterns between countries.
First, material power diﬀerence has been shown to impact voting at the UN Security Council
(Voeten, 2001); while the UN General Assembly does not have the legal power of the UN Security
Council, it would be reasonable to expect that a similar diﬀerence would impact voting at the
UNGA. Thus, I use the Composite Indicator of National Capabilities (Singer et al., 1972, version
5.0) to calculate the diﬀerence between two countries’ power levels; I log-transform this measure
to account for the very high rightward skew. I also include a indications for whether the pair of
countries are either both members of the 5 Permanent Members of the UN Security Council, or
whether one is a P5 member; more precisely, I separate between P3 (US, UK, France) and P2
(Russia and China). Second, I include a binary (0/1) variable for the Post-Cold-War period, deﬁned
as all years after 1991. Third, I include indicators of pre-existing ties between countries, including
alliance obligations (more speciﬁcally, defense pacts, as deﬁned by the Correlates of War Formal
Alliance data v.4.0: see (Gibler, 2009)) and whether or not two countries are jointly Democratic
(Marshall & Jaggers, 2018). Related to pre-existing relationships, I also include the diﬀerences
between countries as expressed by their overall voting patterns, using Bailey et al.’s (2015) ideal
point measures of country positions as measured by voting patterns at the UN General Assembly.
This captures the overall propensity for any two countries to disagree, and also presents a more
diﬃcult test for additional covariates such as speech to explain unique variance not explained by
pre-existing tendencies.

I include several additional variables in the model.29 First, I include year-ﬁxed-eﬀects to account
29In addition to these, as a robustness check, I also estimate models which include a binary (0/1)
variable for whether or not a given country is present in the dyad: for example, the observation for
US/Russia disagreement on Non Proliferation in 1994 will have a 1 for the USA dummy variable;

122

for the overall propensity to disagree in any given year. Second, I include topic-ﬁxed eﬀects to
account for a given topic’s overall propensity to see disagreement, across the entire time frame.
Third, I include a log-transformed measure of the number of words a country spoke on the ﬂoor of
the UN General Assembly; bivariate relationships suggest a curvilinear relationship between the
number of words spoken and the propensity to disagree. While this may be an avenue for future
research, especially given ﬁndings in previous work by the author (Mahmood, 2020b), the reasons
for this curvilinearity are beyond the scope of this paper. However, I do include a quadratic term
for the log-transformed word count in the model.

Finally, given that the research question examines whether speeches at the UNGA are linked to
votes at the UNGA, an important caveat is to emphasize that often, both activities are carried out in
blocs. It is highly common practice at the UNGA for blocs of countries–such as the Arab League,
European Union, Non-Aligned Movement, etc.–to deliver foreign policy speeches as a group,
which by deﬁnition results in an observed similarity in their speech patterns,  
in Equation 4.2.
 
These groups are often formed on the very fact that these countries share a common foreign policy
context, and so they must be accounted for when trying to link foreign policy behavior and speech.
Therefore, during the process of parsing and organizing speeches, I identiﬁed speeches which were
given "on behalf of" an identiﬁable group, or those which are given by a country "in their capacity
as" the chairperson or leader of one of these groups. I include a binary variable for whether two
countries are members of the same group, across all groups for which I have data.

As a matter of theory, however, it is worthwhile to note that this phenomenon complements
the broader argument of this paper: namely, that foreign policy speech is systematically related
to underlying foreign policy interests, and that the political incentives involved in crafting and
delivering a speech are much stronger than previous research has claimed. Indeed, the very fact that
these communities take the time and resources to agree upon and craft joint statements suggests that
these speeches are vehicles for communicating some kind of information relevant to that foreign
policy community. Moreover, a surprisingly common phenomenon is to see countries who are not
a 1 for the Russia dummy variable; and a 0 for all other countries. Some of these models saw
convergence issues given the high degree of freedom.

123

members of a group signing onto statements–such as former Soviet states or aspiring EU states
signing on to statements delivered by the EU. Thus, while I do take these groups into account
empirically, it is noteworthy that the presence of these groups can be seen as an explicit result of
the argument presented in this paper. Future research on the dynamics of foreign policy speech
would do well to examine the politics of crafting these speeches, to see how these politics balance
between sending reinforcing and compensatory (or clarifying) content with respect to the foreign
policies carried out by member countries.

4.4 Results and evaluation

Following the advice of (Brambor et al., 2006), I primarily rely on visualizations as opposed to
interpreting tables alone for the interaction term. Additionally, the linear predictor represents the
change in log-odds of disagreement given a one-unit increase in speech diﬀerence. To facilitate
substantive interpretation, I discuss coeﬃcients in terms of the change in log-odds of disagreement
moving from the 25th percentile to the 75th percentile of disagreement–the Interquartile range
(IQR). In this case, the IQR is about 5 units.

Table 4.2 provides Generalized Linear Model results for a test of Hypothesis 1, where we expect
speech to be less reinforcing as the power diﬀerence between countries increases. The results
support the hypothesis: Figure 4.6 plots the change in the linear predictor for the relationship
of speech to voting (with 95% conﬁdence intervals) across levels of power diﬀerence, with all
other variables held a their median values. Recall again that reinforcing speech should yield a
more positive coeﬃcient for speech; the results are consistent with the expectation that as power
diﬀerence increases, speech becomes more reinforcing. At the lowest levels of CINC diﬀerences,
all else being equal, a one-IQR increase in speech diﬀerence is related to about a 1.3% decrease
in the odds of disagreement. Contrast this with the highest levels of CINC diﬀerence, where a
one-IQR increase in speech diﬀerence is a decrease in the odds of disagreement of around 4%.

Table 4.3 and Table 4.4 provide output for models testing Hypotheses 2 and 3. As noted above,
the expectation for these is that under both conditions–whether Rivalry or joint democracy–we

124

Lin (2)
−20.61
(37.78)
0.06∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.56∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.85∗∗∗
(0.00)
10.96∗∗∗
(0.03)
0.07∗∗∗
(0.02)
−0.10∗∗∗
(0.01)
−0.19
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.01
(0.01)
0.31∗∗∗
(0.00)
−0.35∗∗∗
(0.01)


CINC
−20.10
(33.82)
0.11∗∗∗
(0.00)
−0.02∗∗∗
(0.00)
0.53∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.58∗∗∗
(0.00)
0.80∗∗∗
(0.00)
10.85∗∗∗
(0.03)
0.06∗∗
(0.02)
−0.29∗∗∗
(0.01)
−0.30∗∗
(0.11)
0.11∗∗∗
(0.01)
1.04∗∗∗
(0.05)

−0.00∗∗∗
(0.00)


CINC (2)
−20.57
(37.78)
0.07∗∗∗
(0.00)
−0.01∗∗∗
(0.00)
0.56∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.85∗∗∗
(0.00)
10.97∗∗∗
(0.03)
0.07∗∗∗
(0.02)
−0.10∗∗∗
(0.01)
−0.19
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.00
(0.01)
0.31∗∗∗
(0.00)
−0.35∗∗∗
(0.01)
−0.00∗∗∗
(0.00)


Lin
−20.17
(33.82)
0.10∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.53∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.58∗∗∗
(0.00)
0.80∗∗∗
(0.00)
10.85∗∗∗
(0.03)
0.06∗∗
(0.02)
−0.30∗∗∗
(0.01)
−0.30∗∗
(0.11)
0.11∗∗∗
(0.01)
1.04∗∗∗
(0.05)

(Intercept)

log_cincdiﬀ

SPcur

log_minWC

log_minWC2

jointgroup

UNGAipdiﬀL1

CONTcur

RIVany

oneP2

jointP2

oneP3

jointP3

defense

jointdemL1

proptotalWCgroup

log_cincdiﬀ:SPcur


Year FE
Topic FE
AIC
BIC
Log Likelihood
Deviance
Num. obs.
∗∗∗  < 0.001; ∗∗  < 0.01; ∗  < 0.05

1795443.46
1796182.59

2059296.46
2060019.75

1795408.57
2059465.89
2060176.50
1796160.23
−1029676.95 −897662.73 −1029591.23 −897644.28
1106335.13
1269919.84
2367243
2015035

1269748.40
2367243

1106372.02
2015035

Table 4.2: Binomial model of dyadic disagreement per topic; interaction with power diﬀerence

should see more reinforcing speech (i.e. a more positive slope relating the alignment of speech
and the alignment of votes between any two countries). The results support both expectations.
Figure 4.7 plots the change in linear predictor for the relationship of speech to voting (with 95%
conﬁdence intervals) for Rivalry and non-Rivalry conditions. Among countries with an ongoing
rivalry, a one-IQR increase in speech ideal points is correlated with about a 7.8% increase in the
odds of disagreement. Contrast this with the average for countries not in a rivalry, which is around
an 2.3% decrease in the odds of disagreement (a marginal tendency to hedge).

125

Figure 4.6: Change in linear predictor of Speech diﬀerence vs. Disagreement, across levels of
power diﬀerence (95% conﬁdence intervals)

Figure 4.8 provides a similar plot for the interaction of speech distance and whether two countries
are jointly democratic. Recall the expectation is that jointly Democratic countries will tend to have
more reinforcing speech; the results support this expectation. Among jointly Democratic countries,
a one-IQR change in speech diﬀerence corresponds to around a 1.2% increase in the log odds of
disagreement, versus around a 3% decrease.

Finally, Table 4.5 provides output for a model testing Hypothesis 4. The expectation here
is that as the level of controversy increases, we should expect a corresponding increase in the
reinforcing nature of speech on these issues. Figure 4.9 plots the linear predictor between speech
and disagreement across values of controversy; the results support the expectation that as nations
discuss more controversial issues, their speech tends to be more reinforcing. At the highest levels

126

lllllllllllllllllllllllll−0.02−0.010.00−4−3−2Power differenceMarginal Effect of Speech DiffReinforcing speech versus Power differenceLin (Full)
−20.61
(37.78)
0.06∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.56∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.85∗∗∗
(0.00)
10.96∗∗∗
(0.03)
0.07∗∗∗
(0.02)
−0.10∗∗∗
(0.01)
−0.19
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.01
(0.01)
0.31∗∗∗
(0.00)
−0.35∗∗∗
(0.01)


Riv
−20.74
(33.82)

−0.00∗∗∗
(0.00)
0.57∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.56∗∗∗
(0.00)
0.82∗∗∗
(0.00)
10.84∗∗∗
(0.03)
−0.08∗∗
(0.02)
−0.08∗∗∗
(0.01)
−0.13
(0.11)
0.23∗∗∗
(0.00)
1.22∗∗∗
(0.05)

0.02∗∗∗
(0.00)


Riv (Full)
−20.61
(37.78)
0.06∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.56∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.85∗∗∗
(0.00)
10.96∗∗∗
(0.03)
−0.04
(0.03)
−0.10∗∗∗
(0.01)
−0.21
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.01
(0.01)
0.31∗∗∗
(0.00)
−0.35∗∗∗
(0.01)
0.02∗∗∗
(0.00)


Lin
−20.17
(33.82)
0.10∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.53∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.58∗∗∗
(0.00)
0.80∗∗∗
(0.00)
10.85∗∗∗
(0.03)
0.06∗∗
(0.02)
−0.30∗∗∗
(0.01)
−0.30∗∗
(0.11)
0.11∗∗∗
(0.01)
1.04∗∗∗
(0.05)

(Intercept)

log_cincdiﬀ

SPcur

log_minWC

log_minWC2

jointgroup

UNGAipdiﬀL1

CONTcur

RIVany

oneP2

jointP2

oneP3

jointP3

defense

jointdemL1

proptotalWCgroup

RIVany:SPcur


Year FE
Topic FE
AIC
BIC
Log Likelihood
Deviance
Num. obs.
∗∗∗  < 0.001; ∗∗  < 0.01; ∗  < 0.05

1795443.46
1796182.59

2084485.14
2085196.02

1795391.55
2059465.89
2060176.50
1796143.20
−1029676.95 −897662.73 −1042186.57 −897635.77
1106318.11
1269919.84
2367243
2015035

1290161.08
2379377

1106372.02
2015035

Table 4.3: Binomial model of dyadic disagreement per topic; interaction with Rivalry

of controversy, a one-IQR change in speech diﬀerence is associated with around a 28% increase in
the odds of disagreement, whereas at the lowest level of controversy, a one-IQR change in speech
diﬀerence is associated with a 13% decrease in the odds of disagreement. This is a fairly substantial
change as compared to the other measures; this makes sense, since both controversy levels and
alignment are measured at the UN General Assembly itself, providing a more directly applicable
political environment in which to understand speech.

127

Lin (Full)
−20.61
(37.78)
0.06∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.56∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.85∗∗∗
(0.00)
10.96∗∗∗
(0.03)
0.07∗∗∗
(0.02)
−0.10∗∗∗
(0.01)
−0.19
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.01
(0.01)
0.31∗∗∗
(0.00)
−0.35∗∗∗
(0.01)


Lin
−20.17
(33.82)
0.10∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.53∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.58∗∗∗
(0.00)
0.80∗∗∗
(0.00)
10.85∗∗∗
(0.03)
0.06∗∗
(0.02)
−0.30∗∗∗
(0.01)
−0.30∗∗
(0.11)
0.11∗∗∗
(0.01)
1.04∗∗∗
(0.05)

(Intercept)

log_cincdiﬀ

SPcur

log_minWC

log_minWC2

jointgroup

UNGAipdiﬀL1

CONTcur

RIVany

oneP2

jointP2

oneP3

jointP3

defense

jointdemL1

proptotalWCgroup

jointdemL1:SPcur


Year FE
Topic FE
AIC
BIC
Log Likelihood
Deviance
Num. obs.
∗∗∗  < 0.001; ∗∗  < 0.01; ∗  < 0.05

1795443.46
1796182.59

1811135.93
2059465.89
2060176.50
1811850.19
−1029676.95 −897662.73 −905510.96
1120549.93
1269919.84
2367243
2021867

1106372.02
2015035

JntDem
−21.00
(37.71)

JntDem (Full)

−20.99
(37.77)

−0.00∗∗∗
(0.00)
0.64∗∗∗
(0.01)
−0.08∗∗∗
(0.00)
−0.49∗∗∗
(0.00)
0.88∗∗∗
(0.00)
10.99∗∗∗
(0.03)
0.06∗∗
(0.02)
0.05∗∗∗
(0.01)
−0.04
(0.11)
0.14∗∗∗
(0.00)
0.99∗∗∗
(0.05)

0.29∗∗∗
(0.00)

0.01∗∗∗
(0.00)


−0.01∗∗∗
(0.00)
0.61∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.44∗∗∗
(0.00)
0.86∗∗∗
(0.00)
10.97∗∗∗
(0.03)
0.07∗∗∗
(0.02)
0.03∗∗∗
(0.01)
−0.07
(0.11)
0.15∗∗∗
(0.00)
0.97∗∗∗
(0.05)
0.02∗∗∗
(0.01)
0.29∗∗∗
(0.00)
−0.35∗∗∗
(0.01)
0.01∗∗∗
(0.00)


1808178.39
1808917.72
−904030.20
1117588.39
2021867

Table 4.4: Binomial model of dyadic disagreement per topic; interaction with Joint Democracy

Discussion and conclusion

Out of sample performance is inconclusive. While there is no evidence of overﬁtting–that is,
there is no evidence that the inclusion of speech or the respective interaction terms presented here
cause a decrease in predictive power–there is also no evidence that these additions increase the
out of sample performance of the model. There may be future work to be done on ﬁne-tuning the
measurements of speech alignment as given here–for example, using only speciﬁc parts of speech
or ﬁltering out additional terms, or increasing the dimensionality of the model. While the evidence

128

Figure 4.7: Change in linear predictor of Speech diﬀerence vs. Disagreement, across levels of
rivalry (95% conﬁdence intervals)

presented here is consistent with expectations, the ﬁndings are yet preliminary given the uncertainty
surrounding the measure of speech alignment measurement.

Additionally, these analyses assume a consistent dimensionality over time for speech and issue
space. This may not be the case. For example, Greene et al. (2019) show that the dimensionality
and lexicon of human rights have changed over time; there is no reason to dismiss the possibility
that a similar change has occurred at the UN.30 Moreover, the topic model presented here aggregates
all speeches over time and classiﬁes speeches in topics, independent of time. This is a relatively
30Even if this change has occurred, however, this would not aﬀect the logic of an endogenous
lexicon as provided here, since this change would still be endogenous to the United Nations, and
thus would remain understood across all its member states as the change occurs.

129

ll−0.010.000.010.020.030.00.51.0RivalryMarginal Effect of Speech DiffReinforcing speech versus rivalry (any type)Figure 4.8: Change in linear predictor of Speech diﬀerence vs. Democracy, across levels of joint
Democracy (95% conﬁdence intervals)

coarse method; as others have shown with respect to the UN General Assembly (Bailey et al.,
2015), the agenda of the UNGA has changed over time. Thus, there is room for growth to more
precisely identify the subject matter of an individual speech, paragraph, or sentence, and leverage
this to more precisely measure alignment around a more concretely deﬁned issue.

Finally, dyadic analyses have been subject to strong critique in the past several years (Poast,
2016). While this analysis makes every attempt to account for interdependence–most notably the
inclusion of UN General Assembly voting ideal points, accounting for UN group membership,
and robustness checks with country-level dummy variables–the dyadic structure of the data is
nonetheless a limitation of this paper. Advances in network analysis, for example, would lend
themselves well to further analyzing the relationship between the alignment of behavior and the

130

ll−0.010.000.010.00.51.0Joint DemocracyMarginal Effect of Speech DiffReinforcing speech versus Joint DemocracyLin (Full)
−20.61
(37.78)
0.06∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.56∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.85∗∗∗
(0.00)
10.96∗∗∗
(0.03)
0.07∗∗∗
(0.02)
−0.10∗∗∗
(0.01)
−0.19
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.01
(0.01)
0.31∗∗∗
(0.00)
−0.35∗∗∗
(0.01)


Cont
−20.60
(33.89)

−0.03∗∗∗
(0.00)
0.55∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.55∗∗∗
(0.00)
0.81∗∗∗
(0.00)
10.27∗∗∗
(0.03)
0.05∗∗
(0.02)
−0.09∗∗∗
(0.01)
−0.11
(0.11)
0.23∗∗∗
(0.00)
1.22∗∗∗
(0.05)

0.16∗∗∗
(0.00)


Cont (Full)
−20.47
(37.88)
0.06∗∗∗
(0.00)
−0.03∗∗∗
(0.00)
0.55∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.47∗∗∗
(0.00)
0.84∗∗∗
(0.00)
10.39∗∗∗
(0.04)
0.08∗∗∗
(0.02)
−0.10∗∗∗
(0.01)
−0.20
(0.11)
0.09∗∗∗
(0.01)
0.89∗∗∗
(0.05)
0.00
(0.01)
0.31∗∗∗
(0.00)
−0.33∗∗∗
(0.01)
0.16∗∗∗
(0.00)


Lin
−20.17
(33.82)
0.10∗∗∗
(0.00)
−0.00∗∗∗
(0.00)
0.53∗∗∗
(0.01)
−0.07∗∗∗
(0.00)
−0.58∗∗∗
(0.00)
0.80∗∗∗
(0.00)
10.85∗∗∗
(0.03)
0.06∗∗
(0.02)
−0.30∗∗∗
(0.01)
−0.30∗∗
(0.11)
0.11∗∗∗
(0.01)
1.04∗∗∗
(0.05)

(Intercept)

log_cincdiﬀ

SPcur

log_minWC

log_minWC2

jointgroup

UNGAipdiﬀL1

CONTcur

RIVany

oneP2

jointP2

oneP3

jointP3

defense

jointdemL1

proptotalWCgroup

CONTcur:SPcur


Year FE
Topic FE
AIC
BIC
Log Likelihood
Deviance
Num. obs.
∗∗∗  < 0.001; ∗∗  < 0.01; ∗  < 0.05

1795443.46
1796182.59

2082326.65
2083037.54

1793557.51
2059465.89
2060176.50
1794309.17
−1029676.95 −897662.73 −1041107.33 −896718.76
1104484.07
1269919.84
2367243
2015035

1288002.60
2379377

1106372.02
2015035

Table 4.5: Binomial model of dyadic disagreement per topic; interaction with issue Controversy

alignment of speech; further, these methods would allow for a more rigorous method for modeling
both monadic and dyadic relationships simultaneously (Minhas et al., 2019).

However, the strongest results presented here come from variation in the controversy surround-
ing an issue–as measured by voting disagreement at the UN General Assembly on a given issue. The
most reinforcing speech patterns tend to occur in more controversial issues. There are implications
for future work by scholars studying the United Nations General Assembly, and for the study of
institutions more broadly: interpretation of speeches in relation to other speeches is ideally done in

131

Figure 4.9: Change in linear predictor of Speech diﬀerence vs. Disagreement, across levels of
controversy (95% conﬁdence intervals)

the most controversial issues, since these are the environments where speech will best reﬂect those
underlying patterns of interests. The implicit political ramiﬁcations of hedging on controversial
issues is also a somewhat concerning ﬁnding, since it suggests a positive feedback loop between
leader and political audiences: as an issue becomes more polarized, the costs of de-escalation
become higher–as reﬂected by the fact that despite high controversy, we see countries reinforcing
their already polarized positions as opposed to using speeches to pander towards the other side.

Overall, while there are still additional improvements to be had, this paper has implications for
the study of diplomacy and cooperation, both at the United Nations and in international relations
more broadly. Speciﬁcally, while the United Nations is a vehicle for coordinating foreign policy
(Abbott & Snidal, 1998), and it can reduce foreign policy costs with respect to things like military

132

lllllllllllllllllllllllll−0.0250.0000.0250.0500.00.10.20.30.40.5ControversyMarginal Effect of Speech DiffReinforcing speech versus Controversy (any type)intervention (Thompson, 2009; Chapman, 2011), its role as a vehicle to communicate the alignment
of foreign policy interests more generally has not yet been understood.31 Particularly in highly
politically charged environments–such as during interstate rivalry or in the face of massive power
imbalances–and on the most polarizing, controversial issues in international politics, this role of
institutions in facilitating public diplomacy should be of interest to the scholarly community.

31Chiba & Fang (2014) is an example of work that moves in this direction with respect to the UN.
Other promising work on the topic of speech speciﬁcally includes Ramsay (2011); Bils & Spaniel
(2017).

133

BIBLIOGRAPHY

134

BIBLIOGRAPHY

Abbott, Kenneth W & Duncan Snidal. 1998. Why states act through formal international organi-

zations. Journal of conﬂict resolution 42(1). 3–32.

ABBYY Production LLC. 2013. Abbyy ﬁnereader pdf. Accessed online: https://pdf.abbyy.com/.

Aitchison, John. 1986. The statistical analysis of compositional data .
Alker, Hayward R. 1964. Dimensions of conﬂict in the general assembly. American Political

Science Review 58(03). 642–657.

Alker, Hayward R. Jr. & Bruce M. Russett. 1965. World politics in the general assembly. Yale

University Press.

Allen, Susan Hannah & Amy T Yuen. 2020. Action or inaction: United nations security council

activity, 1994–2013. Journal of Peace Research 57(5). 658–665.

Andersson, Andreas. 2002. United nations intervention by united democracies? state commitment

to un interventions 1991-99. Cooperation and Conﬂict 37(4). 363–386.

Bailey, Michael A, Anton Strezhnev & Erik Voeten. 2015. Estimating dynamic state preferences

from united nations voting data. Journal of Conﬂict Resolution 0022002715595700.

Ball, M Margaret. 1951. Bloc voting in the general assembly. International Organization 5(1).

3–31.

Barnett, Michael & Martha Finnemore. 2004. Rules for the world: International organizations in

global politics. Cornell University Press.

Baturo, Alexander & Niheer Dasandi. 2017. What drives the international development agenda?
an nlp analysis of the united nations general debate 1970–2016. In 2017 international conference
on the frontiers and advances in data science (fads), 171–176. IEEE.

Baturo, Alexander, Niheer Dasandi & Slava Mikhaylov. 2017. Understanding state preferences

with text as data: Introducing the un general debate corpus. Research & Politics .

Baum, Matthew A. 2004. Going private: Public opinion, presidential rhetoric, and the domestic
politics of audience costs in us foreign policy crises. Journal of Conﬂict Resolution 48(5).
603–631.

Baum, Matthew A & Tim J Groeling. 2009. War stories: The causes and consequences of public

views of war. Princeton University Press.

Baum, Matthew A & Philip BK Potter. 2008. The relationships between mass media, public

opinion, and foreign policy: Toward a theoretical synthesis. Annu. Rev. Polit. Sci. 11. 39–65.

135

Beardsley, Kyle. 2013. The un at the peacemaking-peacebuilding nexus. Conﬂict Management and

Peace Science 0738894213491354.

Beardsley, Kyle & Holger Schmidt. 2012. Following the ﬂag or following the charter? examining
the determinants of un involvement in international crises, 1945–2002. International Studies
Quarterly 56(1). 33–49.

Bellamy, Alex J & Paul D Williams. 2013. Providing peacekeepers: the politics, challenges, and

future of united nations peacekeeping contributions. OUP Oxford.

Berry, William D, Matt Golder & Daniel Milton. 2012.

interaction. The Journal of Politics 74(3). 653–671.

Improving tests of theories positing

Bils, Peter & William Spaniel. 2017. Policy bargaining and militarized conﬂict. Journal of

Theoretical Politics 29(4). 647–678.

Binder, Martin. 2015. Paths to intervention what explains the uns selective response to humanitarian

crises? Journal of Peace Research 52(6). 712–726.

Binder, Martin & Jonathan Golub. 2020. Civil Conﬂict and Agenda-Setting Speed in the United

Nations Security Council. International Studies Quarterly 64(2). 419–430.

Binder, Martin & Monika Heupel. 2015. The legitimacy of the un security council: Evidence from

recent general assembly debates. International Studies Quarterly 59(2). 238–250.

Blei, David M & John D Laﬀerty. 2006. Dynamic topic models.

international conference on machine learning, 113–120. ACM.

In Proceedings of the 23rd

Blei, David M, Andrew Y Ng & Michael I Jordan. 2003. Latent dirichlet allocation. the Journal

of machine Learning research 3. 993–1022.

Boese, Vanessa A. 2019. How (not) to measure democracy. International Area Studies Review

22(2). 95–127.

Boix, Carles. 2011. Democracy, development, and the international system. American political

science review 809–828.

Bove, Vincenzo & Leandro Elia. 2011. Supplying peace: Participation in and troop contribution

to peacekeeping missions. Journal of Peace Research 48(6). 699–714.

Brambor, Thomas, William Roberts Clark & Matt Golder. 2006. Understanding interaction models:

Improving empirical analyses. Political analysis 14(1). 63–82.

Breheny, Patrick & Woodrow Burchett. 2015. visreg: Visualization of regression models. R

package version 2.2-0.

Budge, Ian, David Robertson & Derek Hearl. 1987. Ideology, strategy and party change: spatial

analyses of post-war election programmes in 19 democracies. Cambridge University Press.

Carter, David B & Randall W Stone. 2015. Democracy and multilateralism: the case of vote buying

in the un general assembly. International Organization 69(1). 1–33.

136

Chapman, Terrence L. 2011. Securing approval: domestic politics and multilateral authorization

for war. University of Chicago Press.

Cheibub, José Antonio, Jennifer Gandhi & James Raymond Vreeland. 2010. Democracy and

dictatorship revisited. Public choice 143(1-2). 67–101.

Chelotti, Nicola, Niheer Dasandi & Slava Jankin Mikhaylov. 2018. Do intergovernmental organi-
zations have a socialization eﬀect on member state preferences? evidence from the un general
debate. Working paper (accessed via SocArXiv November 2020) .

Chiba, Daina & Songying Fang. 2014. Institutional opposition, regime accountability, and interna-

tional conﬂict. The Journal of Politics 76(3). 798–813.

Claude, Inis L. 1966. Collective legitimization as a political function of the united nations.

International Organization 20(03). 367–379.

Colaresi, Michael P. 2005. Scare tactics: The politics of international rivalry. Syracuse University

Press.

Colaresi, Michael P. 2014. Democracy declassiﬁed: The secrecy dilemma in national security.

Oxford University Press.

Colaresi, Michael P, Karen Rasler & William R Thompson. 2008. Strategic rivalries in world

politics: position, space and conﬂict escalation. Cambridge University Press.

Collier, Kenneth. 2018. Speechwriting in the institutionalized presidency: Whose line is it?

Lexington Books.

Collobert, Ronan, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu & Pavel Kuksa.
2011. Natural language processing (almost) from scratch. Journal of machine learning research
12. 2493–2537.

Crescenzi, Mark JC. 2007. Reputation and interstate conﬂict. American Journal of Political Science

51(2). 382–396.

Diamond, Larry Jay. 1996. Is the third wave over? Journal of democracy 7(3). 20–37.
Dĳkhuizen, Frederieke & Michal Onderco. 2019. Sponsorship behaviour of the brics in the united

nations general assembly. Third World Quarterly 1–16.

Dreher, Axel, Peter Nunnenkamp & Rainer Thiele. 2008. Does us aid buy un general assembly

votes? a disaggregated analysis. Public Choice 136(1-2). 139–164.

Elﬀ, Martin. 2013. A dynamic state-space model of coded political texts. Political Analysis 21(2).

217–232.

Fariss, Christopher J. 2014. Respect for human rights has improved over time: Modeling the

changing standard of accountability. American Political Science Review 108(2). 297–318.

Farrell, Joseph & Robert Gibbons. 1989a. Cheap talk can matter in bargaining. Journal of Economic

Theory 48(1). 221–237.

137

Farrell, Joseph & Robert Gibbons. 1989b. Cheap talk with two audiences. The American Economic

Review 79(5). 1214–1223.

Fearon, James D. 1994. Domestic political audiences and the escalation of international disputes.

American Political Science Review 88(03). 577–592.

Fearon, James D. 1995. Rationalist explanations for war.

379–414.

International organization 49(03).

Findley, Michael G & Tze Kwang Teo. 2006. Rethinking third-party interventions into civil wars:

An actor-centric approach. Journal of Politics 68(4). 828–837.

Food & Agriculture Organization of the United Nations. 2019a. Annual population. Accessed

online: http://www.fao.org/faostat/en.

Food & Agriculture Organization of the United Nations. 2019b. Gross domestic product. Accessed

online: http://www.fao.org/faostat/en.

Fortna, Virginia Page. 2008. Does peacekeeping work?: shaping belligerents’ choices after civil

war. Princeton University Press.

Fuhrmann, Matthew & Benjamin Tkach. 2015. Almost nuclear: introducing the nuclear latency

dataset. Conﬂict Management and Peace Science 32(4). 443–461.

Gartzke, Erik. 1999. War is in the error term. International Organization 567–587.
Gartzke, Erik & Dong-Joon Jo. 2006. The aﬃnity of nations index, 1946-2002. New York, NY:

Columbia University .

Gartzke, Erik & Yonatan Lupu. 2012. Still looking for audience costs. Security Studies 21(3).

391–397.

Geddes, Barbara, Joseph Wright & Erica Frantz. 2014. Autocratic breakdown and regime transi-

tions: A new data set. Perspectives on Politics 12(2). 313–331.

Gent, Stephen E. 2007. Strange bedfellows: The strategic dynamics of major power military

interventions. Journal of Politics 69(4). 1089–1102.

Gent, Stephen E. 2008. Going in when it counts: Military intervention and the outcome of civil

conﬂicts*. International Studies Quarterly 52(4). 713–735.

Gibler, D.M. 2009. International military alliances, 1648-2008. Congressional Quarterly Press.

ISBN:978-1-56802-824-8.

Gilligan, Michael & Stephen J Stedman. 2003. Where do the peacekeepers go?

Studies Review 5(4). 37–54.

International

Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg & Håvard Strand.

2002. Armed conﬂict 1946-2001: A new dataset. Journal of peace research 39(5). 615–637.

138

Greene, Kevin T, Baekkwan Park & Michael Colaresi. 2019. Machine learning human rights and
wrongs: How the successes and failures of supervised learning algorithms can inform the debate
about information eﬀects. Political Analysis 27(2). 223–230.

Grimmer, Justin & Brandon M Stewart. 2013. Text as data: The promise and pitfalls of automatic

content analysis methods for political texts. Political Analysis 21(3). 267–297.

Guisinger, Alexandra & Alastair Smith. 2002. Honest threats: The interaction of reputation and

political institutions in international crises. Journal of Conﬂict Resolution 46(2). 175–200.

Gurciullo, Stefano & Slava J Mikhaylov. 2017. Detecting policy preferences and dynamics in the un
general debate with neural word embeddings. In 2017 international conference on the frontiers
and advances in data science (fads), 74–79. IEEE.

Habermas, Jürgen. 1984. The theory of communicative action, volume i. Beacon.
Hayes, Danny & Matt Guardino. 2013. Inﬂuence from abroad: foreign voices, the media, and us

public opinion. Cambridge University Press.

Hecht, Catherine. 2016. The shifting salience of democratic governance: Evidence from the united

nations general assembly general debates. Review of International Studies 42(5). 915–938.

Hobbes, Thomas & Crawford Brough Macpherson. 1968. Leviathan; edited with an introduction

by cb macpherson. Penguin.

House, Freedom. 2014. Freedom in the world. Http://www.freedomhouse.org.
Jacobsen, Kurt. 1969. Sponsorships in the united nations: A system analysis. Journal of Peace

Research 6(3). 235–256.

Jervis, Robert. 1970. The logic of images. Princeton, NJ: Princeton UP .
Johnstone, Ian. 2003. Security council deliberations: the power of the better argument. European

Journal of International Law 14(3). 437–480.

Johnstone, Ian. 2005. Discursive power in the un security council. J. Int’l L & Int’l Rel. 2. 73.
Kam, Cindy & Robert J Franzese. 2007. Modeling and interpreting interactive hypotheses in

regression analysis. University of Michigan Press.

Kashin, Konstantin. 2014. panelar: Estimation of linear ar(1) panel data models with cross-

sectional heteroskedasticity and/or correlation. R package version 0.1.

Kathman, Jacob D. 2013. United nations peacekeeping personnel commitments, 1990–2011.

Conﬂict Management and Peace Science 30(5). 532–549.

Kathman, Jacob D & Molly M Melin. 2017. Who keeps the peace? understanding state contributions

to un peacekeeping operations. International Studies Quarterly 61(1). 150–162.

Katzman, Kenneth. 2019. Bahrain: Unrest, security, and u.s. policy.

139

Keohane, Robert O. 1969. Institutionalization in the united nations general assembly. International

Organization 23(4). 859–896.

Keohane, Robert O. 2012. Twenty years of institutional liberalism. International Relations 26(2).

125–138.

Keohane, Robert O & Lisa L Martin. 1995. The promise of institutionalist theory. International

security 20(1). 39–51.

Kertzer, Joshua D. 2016. Resolve in international politics. Princeton University Press.
Kertzer, Joshua D & Ryan Brutger. 2016. Decomposing audience costs: Bringing the audience

back into audience cost theory. American Journal of Political Science 60(1). 234–249.

Kim, Soo Yeon & Bruce Russett. 1996. The new politics of voting alignments in the united nations

general assembly. International Organization 50(04). 629–652.

Krain, Matthew. 1998. Contemporary democracies revisited: Democracy, political violence, and

event count models. Comparative Political Studies 31(2). 139–164.

Krebs, Ronald R & Patrick Thaddeus Jackson. 2007. Twisting tongues and twisting arms: The

power of political rhetoric. European Journal of International Relations 13(1). 35–66.

Kurizaki, Shuhei. 2007. Eﬃcient secrecy: Public versus private threats in crisis diplomacy.

American Political Science Review 101(03). 543–558.

Kurizaki, Shuhei & Taehee Whang. 2015. Detecting audience costs in international disputes.

International Organization 69(4). 949–980.

Kydd, Andrew. 1997. Sheep in sheep’s clothing: Why security seekers do not ﬁght each other.

Security Studies 7(1). 114–155.

Kydd, Andrew. 2005. Trust and mistrust in international relations. Princeton University Press.
Lake, David A. 2009. Hierarchies in international relations. Cornell University Press.
Laver, Michael, Kenneth Benoit & John Garry. 2003. Extracting policy positions from political

texts using words as data. American Political Science Review 97(02). 311–331.

Leifeld, Philip. 2013. texreg: Conversion of statistical model output in R to LATEX and HTML tables.
Journal of Statistical Software 55(8). 1–24. Accessed online: http://www.jstatsoft.org/v55/i08/.
Lektzian, David & Mark Souva. 2007. An institutional theory of sanctions onset and success.

Journal of Conﬂict Resolution 51(6). 848–871.

Lemke, Douglas & Patrick M Regan. 2004. Interventions as inﬂuence. The Scourge of War: New

extensions on an old problem 145–168.

Levendusky, Matthew S & Michael C Horowitz. 2012. When backing down is the right decision:

Partisanship, new information, and audience costs. The Journal of Politics 74(02). 323–338.

140

Levitsky, Steven & Lucan Way. 2005.

democracy 16(3). 20–34.

International linkage and democratization. Journal of

Levitsky, Steven & Lucan A Way. 2010. Competitive authoritarianism: Hybrid regimes after the

cold war. Cambridge University Press.

Li, Quan. 2006. Democracy, autocracy, and tax incentives to foreign direct investors: A cross-

national analysis. The Journal of Politics 68(1). 62–74.

Liao, Tim Futing. 1994.

linear models. Sage.

Interpreting probability models: Logit, probit, and other generalized

Lĳphart, Arend. 1963. The analysis of bloc voting in the general assembly: A critique and a

proposal. American Political Science Review 57(4). 902–917.

Lindberg, Staﬀan I, Michael Coppedge, John Gerring & Jan Teorell. 2014. V-dem: A new way to

measure democracy. Journal of Democracy 25(3). 159–169.

Long, Scott J. 1997. Regression models for categorical and limited dependent variables. Sage.
Lührmann, Anna & Staﬀan I Lindberg. 2019. A third wave of autocratization is here: what is new

about it? Democratization 26(7). 1095–1113.

Mahmood, Zuhaib. 2020a. Compositional speciﬁcity and diﬀuseness of un general assembly speech

across world issues. Dissertation Chapter .

Mahmood, Zuhaib. 2020b. Material capacity and political incentives as drivers of un general

assembly speeches. Dissertation Chapter .

Mahmood, Zuhaib & Michael P. Colaresi. 2016. Tracking state preferences in high resolution using
un speeches. Paper prepared for the Annual meeting of the International Studies Association .
Mahmood, Zuhaib & Michael P. Colaresi. 2017. Using un general assembly speeches to track state
interests and the dimensionality of world politics. Paper prepared for the Annual meeting of the
International Studies Association .

Maier, Marco J. 2014. Dirichletreg: Dirichlet regression for compositional data in r Accessed

online: http://epub.wu.ac.at/4077.

Marshall, Monty G & Keith Jaggers. 2018. Polity iv project: Political regime characteristics and

transitions, 1800-2017 .

Mearsheimer, John J. 1994. The false promise of international institutions. International security

19(3). 5–49.

Mearsheimer, John J. 1995. A realist reply. International Security 20(1). 82–93.
Mearsheimer, John J. 2001. The tragedy of great power politics. WW Norton & Company.
Mechkova, Valeriya, Anna Lührmann & Staﬀan I Lindberg. 2017. How much democratic back-

sliding? Journal of Democracy 28(4). 162–169.

141

Medzihorsky, Juraj, Milos Popovic & Erin K Jenne. 2017. Rhetoric of civil conﬂict management:
United nations security council debates over the syrian civil war. Research & Politics 4(2).
2053168017702982.

Mimno, David & Moontae Lee. 2014. Low-dimensional embeddings for interpretable anchor-based
topic inference. In Proceedings of the 2014 conference on empirical methods in natural language
processing (emnlp), 1319–1328.

Minhas, Shahryar, Peter D Hoﬀ & Michael D Ward. 2019.

Inferential approaches for network

analysis: Amen for latent factor models. Political Analysis 27(2). 208–222.

Monroe, Burt L, Michael P Colaresi & Kevin M Quinn. 2008. Fightin’words: Lexical feature
selection and evaluation for identifying the content of political conﬂict. Political Analysis 16(4).
372–403.

Morrow, James D. 1986. A spatial model of international conﬂict. The American Political Science

Review 1131–1150.

Morrow, James D. 1994. Alliances, credibility, and peacetime costs. Journal of Conﬂict Resolution

38(2). 270–297.

Morrow, James D. 2000. Alliances: Why write them down? Annual Review of Political Science

3(1). 63–83.

Nations, The United. 1945. Charter of the united nations. 1 UNTS XVI. Accessed August 2020.

Https://www.un.org/en/charter-united-nations/index.html.

Neack, Laura. 1995. Un peace-keeping: In the interest of community or self? Journal of Peace

Research 32(2). 181–196.

Newcombe, Hanna, Michael Ross & Alan G Newcombe. 1970. United nations voting patterns.

International Organization 24(1). 100–121.

Pemstein, Daniel, Stephen A Meserve & James Melton. 2010. Democratic compromise: A latent

variable analysis of ten measures of regime type. Political Analysis 18(4). 426–449.

Pettersson, Therése, Stina Högbladh & Magnus Öberg. 2019. Organized violence, 1989–2018 and

peace agreements. Journal of Peace Research 56(4). 589–603.

Pevehouse, Jon, Timothy Nordstrom & Kevin Warnke. 2004. The correlates of war 2 international
governmental organizations data version 2.0. Conﬂict Management and Peace Science 21(2).
101–119.

Poast, Paul. 2016. Dyads are dead, long live dyads! the limits of dyadic designs in international

relations research. International Studies Quarterly 60(2). 369–374.

Press, Daryl Grayson. 2005. Calculating credibility: How leaders assess military threats. Cornell

University Press.

142

Proksch, Sven-Oliver & Jonathan B Slapin. 2012. Institutional foundations of legislative speech.

American Journal of Political Science 56(3). 520–537.

Przeworski, Adam. 2000. Democracy and development: political institutions and well-being in the

world, 1950-1990, vol. 3. Cambridge University Press.

Quinn, Kevin M, Burt L Monroe, Michael Colaresi, Michael H Crespin & Dragomir R Radev.
2010. How to analyze political attention with minimal assumptions and costs. American Journal
of Political Science 54(1). 209–228.

Rai, Kul B. 1980. Foreign aid and voting in the un general assembly, 1967—1976. Journal of

Peace Research 17(3). 269–277.

Ramsay, Kristopher W. 2004. Politics at the waters edge: Crisis bargaining and electoral competi-

tion. Journal of Conﬂict Resolution 48(4). 459–486.

Ramsay, Kristopher W. 2011. Cheap talk diplomacy, voluntary negotiations, and variable bargaining

power. International Studies Quarterly 55(4). 1003–1023.

Recchia, Stefano. 2015. Did chirac say ‘non’? revisiting un diplomacy on iraq, 2002-03. Political

Science Quarterly 130(4). 625–654.

Richardson, Leonard. 2007. Beautiful soup documentation. Python library .
Roberts, Margaret E, Brandon M Stewart, Dustin Tingley et al. 2014. stm: R package for structural

topic models. Journal of Statistical Software 10(2). 1–40.

Ruggie, John Gerard. 1982. International regimes, transactions, and change: embedded liberalism

in the postwar economic order. International organization 36(2). 379–415.

Russett, Bruce M. 1966. Discovering voting groups in the united nations. American Political

Science Review 60(02). 327–339.

Sartori, Anne E. 2002. The might of the pen: A reputational theory of communication in interna-

tional disputes. International Organization 56(1). 121–149.

Sartori, Anne E. 2005. Deterrence by diplomacy. Princeton University Press.
Schelling, Thomas C. 1960. The strategy of conﬂict. Cambridge, Mass .
Schultz, Kenneth A. 1998. Domestic opposition and signaling in international crises. American

Political Science Review 92(04). 829–844.

Schultz, Kenneth A. 2001. Democracy and coercive diplomacy, vol. 76. Cambridge University

Press.

Schultz, Kenneth A. 2012. Why we needed audience costs and what we need now. Security Studies

21(3). 369–375.

Schweller, Randall L. 1996. Neorealism’s status-quo bias: What security dilemma?

Studies 5(3). 90–121.

Security

143

Singer, J David. 1963. Inter-nation inﬂuence: a formal model. American Political Science Review

57(02). 420–430.

Singer, J David, Stuart Bremer & John Stuckey. 1972. Capability distribution, uncertainty, and

major power war, 1820-1965. Peace, war, and numbers 19.

Slantchev, Branislav L. 2012. Audience cost theory and its audiences. Security Studies 21(3).

376–382.

Slapin, Jonathan B & Sven-Oliver Proksch. 2008. A scaling model for estimating time-series party

positions from texts. American Journal of Political Science 52(3). 705–722.

Smith, Courtney B. 2006. Politics and process at the united nations: the global dance. Lynne

Rienner Boulder, CO.

Snyder, Jack & Erica D Borghard. 2011. The cost of empty threats: A penny, not a pound. American

Political Science Review 105(3). 437–456.

South, Andy. 2011. rworldmap: A new r package for mapping global data. The R Journal 3(1).

35–43.

Stasavage, David. 2004. Open-door or closed-door? transparency in domestic and international

bargaining. International Organization 58. 667–704.

Steele, Brent J. 2007. Making words matter: The asian tsunami, darfur, and “reﬂexive discourse”

in international politics. International Studies Quarterly 51(4). 901–925.

Steﬀek, Jens. 2003. The legitimation of international governance: A discourse approach. European

Journal of International Relations 9(2). 249–275.

Svolik, Milan W. 2012. The politics of authoritarian rule. Cambridge University Press.
Tarar, Ahmer & Bahar Leventoğlu. 2009. Public commitment in crisis bargaining. International

Studies Quarterly 53(3). 817–839.

Tarar, Ahmer & Bahar Leventoğlu. 2013. Limited audience costs in international crises. Journal

of Conﬂict Resolution 57(6). 1065–1089.

Terechshenko, Zhanna. 2020. Hot under the collar: A latent measure of interstate hostility. Journal

of Peace Research .

Thompson, Alexander. 2009. Channels of power: The un security council and us statecraft in iraq.

Cornell University Press.

Thompson, William & David Dreyer. 2011. Handbook of international rivalries. CQ Press.
Thyne, Clayton L. 2009. How international relations aﬀect civil conﬂict: Cheap signals, costly

consequences. Rowman & Littleﬁeld.

Tomz, Michael. 2007. Domestic audience costs in international relations: An experimental ap-

proach. International Organization 61(4). 821–840.

144

Tomz, Michael R & Jessica LP Weeks. 2013. Public opinion and the democratic peace. American

political science review 107(4). 849–865.

Trachtenberg, Marc. 2012. Audience costs: An historical analysis. Security Studies 21(1). 3–42.
Trager, Robert F. 2011. Multidimensional diplomacy. International Organization 65(03). 469–506.
Trager, Robert F. 2013. How the scope of a demand conveys resolve. International Theory 5(3).

414–445.

Trager, Robert F. 2017. Diplomacy: Communication and the origins of international order.

Cambridge University Press.

Trager, Robert F & Lynn Vavreck. 2011. The political costs of crisis bargaining: Presidential

rhetoric and the role of party. American Journal of Political Science 55(3). 526–545.

Treier, Shawn & Simon Jackman. 2008. Democracy as a latent variable. American Journal of

Political Science 52(1). 201–217.

Uzonyi, Gary. 2015. Refugee ﬂows and state contributions to post-cold war un peacekeeping

missions. Journal of Peace Research 52(6). 743–757.

Venables, W. N. & B. D. Ripley. 2002. Modern applied statistics with s. New York: Springer 4th

edn. ISBN 0-387-95457-0.

Venables, William N & Brian D Ripley. 2013. Modern applied statistics with s-plus. Springer

Science & Business Media.

Vincent, Jack E. 1971. Predicting voting patterns in the general assembly. American Political

Science Review 65(02). 471–498.

Vincent, Jack E. 1972. An application of attribute theory to general assembly voting patterns, and

some implications. International Organization 26(03). 551–582.

Voeten, Erik. 2000. Clashes in the assembly. International organization 185–215.
Voeten, Erik. 2001. Outside options and the logic of security council action. American Political

Science Review 95(4). 845–858.

Voeten, Erik. 2005. The political origins of the un security council’s ability to legitimize the use

of force. International Organization 59(03). 527–557.

Vreeland, James Raymond & Axel Dreher. 2014. The political economy of the united nations

security council: money and inﬂuence. Cambridge University Press.

Waheed, Moniza, Andreas Schuck, Peter Neĳens & Claes de Vreese. 2013. Values in un speeches:
International Communication Gazette 75(7).

Understanding the speechwriters’ perspectives.
597–618.

Waltz, Kenneth. 1979. Theory of international politics. McGraw-Hill.

145

Waltz, Kenneth Neal. 1959. The united nations as a political institution. Columbia University

Press.

Wang, Te-Yu. 1999. Us foreign aid and un voting: An analysis of important issues. International

Studies Quarterly 43(1). 199–210.

Weber, Max. 1978. Economy and society. Univ of California Press.
Weeks, Jessica L. 2008. Autocratic audience costs: Regime type and signaling resolve. International

Organization 62(01). 35–64.

Weeks, Jessica L. 2012. Strongmen and straw men: Authoritarian regimes and the initiation of

international conﬂict. American Political Science Review 106(2). 326–347.

Western, Bruce. 1995. Concepts and suggestions for robust regression analysis. American Journal

of Political Science 786–817.

White, Jonathan & Lea Ypi. 2011. On partisan political justiﬁcation. American political science

review 381–396.

Zeileis, Achim. 2004. Econometric computing with HC and HAC covariance matrix estimators.

Journal of Statistical Software 11(10). 1–17. doi:10.18637/jss.v011.i10.

Zeileis, Achim. 2006. Object-oriented computation of sandwich estimators. Journal of Statistical

Software 16(9). 1–16. doi:10.18637/jss.v016.i09.

Zeileis, Achim, Francisco Cribari-Neto, Bettina Gruen, Ioannis Kosmidis, Alexandre B Simas,

Andrea V Rocha & Maintainer Achim Zeileis. 2016. Package ‘betareg’. R package .

146