“united we stand, divided we fall.”which countries join coalitions more often in gatt/wto

negotiations?

Gabriel Cepaluni†

Ivan F. de A. Lopes Fernandes‡

Julio A. Z. Trecenti§

Athos P. Damiani¶

This version: June 14, 2014

First version: January 14, 2013

Preliminary draft

We are grateful to Silvia Lopes de Paula Ferrari who statistically advised us in earlier stages of the project.

Matthew Winters careful read an earlier versions of the manuscript, suggesting several improvements.

Thomas Sattler shared his data on the number of delegates in Geneva. Robert Wolfe, Rodolpho Bern-

abel, Umberto Mignozzetti and Feliciano Guimaraes provide us with constructive comments. We also thank

several participants at two seminars at the Department of Statistics from the University of Sao Paulo. All

remaining errors are our own.†Assistant Professor, Department of International Relations, Sao Paulo State University.‡PhD Candidate, Department of Political Science, University of Sao Paulo.§Researcher, Department of Statistics, University of Sao Paulo.¶Researcher, Department of Statistics, University of Sao Paulo.

Coalition-formation is an important tool to leverage countries’ bargaining power in the

GATT/WTO negotiations. Unlike the “weapons of the weak” reasoning, several econo-

metric models and specifications show that larger economies have a higher probability of

joining coalitions. Challenging the view that middle powers have a distinguishable collec-

tivist behavior, even non-linear models show that the relationship between real GDP and

coalition entry has an almost linear shape. Large economies join coalitions more because

they are more equipped to pay transaction costs and more prepared to deal with the risks

of uncertain negotiations. Countries more open to trade also join coalitions often—a less

surprising result since the GATT/WTO is a pro-open trade international institution. At

last, unlike the “democratic peace” literature, we do not find that democracies cooperate

more than dictatorships. When dictatorships become more democratic, they tend to join

coalitions more up to a threshold; after that, the effect decreases forming an inverted U-

shaped curve, suggesting that the correlation between political regime and cooperation is

not straightforward.

I. Introduction

Few studies have explored the role of international bargaining coalitions (Hamilton and

Whalley, 1989; Higgott and Cooper, 1990; Narlikar and Tussie, 2004; Costantini et al., 2007;

Narlikar, 2003) and even fewer have done that quantitatively (Costantini et al., 2007). In

this paper, we propose a still unanswered basic question: “which are the most important

economic and political factors that increase the probability of countries join coalitions?”

The question is important because some authors take for granted that coalitions are one of

the few tools that weak countries have at their disposal (Drahos, 2003; Braithwaite, 2004).

As Narlikar points out eloquently: “United we stand, divided we fall. The reasoning of the

weak in their dealings with the strong is simple and direct” (Narlikar, 2003, p. 1).

Some practitioners and scholars believe that developing countries have a voice against

the power of the developed countries at the GATT/WTO; then, it is better to negotiate

multilaterally than bilaterally.1 In particular, GATT/WTO coalitions are considered an

effective tool at the disposal of developing countries. Whether a useful tool or not, we

found that powerful countries take advantage of this tool more often. We explain our

finding by arguing that collectively negotiating at the GATT/WTO is costly. Among other

costs, countries have to train and keep diplomatic missions in Geneva. We tested the

proposed causal mechanism and found evidences that the number of delegates in Geneva

is an mediator variable between real GDP and coalition entry.

Our paper employs an unique country-level panel data set. It contains 3,189 observa-

tions of 144 countries bargaining at the GATT/WTO from 1982 to 2008. Our main result

contradicts the view that coalitions are “weapons of the weak” (Drahos, 2003; Braithwaite,

2004). The effect of GDP is large—an increase of one standard deviation increases the

likelihood of a country joining a coalition by 60%. When we used 7 years lag of natu-

ral disaster as an instrumental variable for real GDP, the results get even larger—around

145%. We also test the argument that coalitions at the GATT/WTO are weapons used

more often by middle powers. The main assumption of the middle power literature is

that weak countries are irrelevant at international negotiations because they have scarce

resources, whereas powerful countries are capable of acting unilaterally and do not want

1The former Brazilian Minister of Foreign Affairs and Brazilian Ambassador in the WTO, Celso Lafer,wrote a book defending the importance of the WTO to Brazil (Lafer, 1998). He argues that the WTO offersmore opportunities to middle income countries than bilateral negotiations with powerful states. Currently,the Secretary-General of the WTO is a Brazilian diplomat. See also (Narlikar, 2003). (Steinberg, 2002)provides a different take on the subject.

1

to bind themselves with coalitions commitments. Middle powers are, then, prone to have

a collectivist behavior (Cooper, Higgott and Nossal, 1993; Higgott and Cooper, 1990). To

test the middle power hypothesis, we examine the functional form of the relationship be-

tween GDP and coalition entry. Theoretically, the middle power hypothesis has a quadratic

functional form, whereas the weapons of the weak notion has a decreasing linear function.

We employ Generalized Additive Models (GAM) to graphically display the non-linear re-

lationship between our dependent and independent variables. The positive relationship

between real GDP and coalition entry is quasi-linear, even using non-linear models.2 Other

findings concern the relationship between coalition entry and both trade openness and po-

litical regimes. Countries more open to trade are also more likely to join coalitions. Not

strikingly, considering that the GATT/WTO is a pro-open trade institution. Finally, con-

trary to the “democratic peace” literature (Olson, 1965; Chamberlin, 1974; Hardin, 1982;

Ostrom, 1990; Esteban, 2001), democracies do not cooperate more by means of coalition

building at the GATT/WTO negotiations. As it should be clear, GATT/WTO members

are already cooperating by joining the international trade system.

Another contribution of our study is the measurement of our dependent variable—

coalition entry. Most studies have only analyzed few cases of bargaining coalitions at the

GATT/WTO (Narlikar, 2003; Hamilton and Whalley, 1989; Higgott and Cooper, 1990;

Narlikar and Tussie, 2004; Odell, 2009). As far as we know, a measure of coalition en-

try that enables quantitative analyzes has not been used before this research. Our main

source of information is the WTO website.3 The section “Groups in the Negotiations” have

information on groups participating in the WTO negotiations, such as: coalition names;

country member names; the issues they are negotiating; coalitions’ websites; and the na-

ture of coalitions (e.g. custom union, regional, or broad interests). We have supplemented

the WTO data by reading the specialized literature.4 The main information required to

build our dependent variable is coalition names, their members, and the year of coalition

formation—see Table 1. Coalitions’ first year is our main benchmark to acknowledge the

number and names of countries that join a coalition. We choose the year of the coalition

formation as benchmark because it is easier to observe which are the members of the coali-

tion at its onset. Many coalitions are informal groups, but members present a position

2We cannot include time-invariant variables in panel fixed-effects models—such as the North-South hemi-spheres. But we include it in other models showed in this paper.

3http://www.wto.org/.4Important works for supplementing the WTO website are: Singh (2006), Narlikar (2003), Patel (2007),

Yu III (2008) and Wolfe (2006).

2

paper signing their names before the negotiations or even maintain a website. They rarely,

however, go public stating they are leaving a coalition. The unit of analysis of our data set

is at the country-level. We observed how many times a country i has entered in trade coali-

tions at year t. As important as it is, we cannot distinguish between long- and short-lived

coalitions with our data; countries only join once each coalition.

In the next section, we review the history of the GATT/WTO, focusing on the increasing

role of bargaining coalitions. In the third section, we discuss our theoretical argument

proposing an explanation for our results. In the fourth section, we describe our data and

variables. In the fifth section, we discuss our methodology and our main results. Finally,

we conclude our paper.

II. Overall Context

In recent years, bargaining coalitions at the WTO has become increasingly visible.

While there had been few coalitions in GATT’s early years, the number of coalitions in-

creased substantially since the 1980s (Narlikar, 2003, p. 34). GATT was a “Rich Man’s

Club,” where tariff protection and US domination in terms of international trade were the

rule. At the formation of the GATT, 11 of the 23 founding members were developing

countries. Despite their relatively large proportion, developing countries maintained a low

profile in the GATT (Narlikar, 2003, p. 35). According to Rubens Ricupero—former Brazil-

ian ambassador in Geneva (1987-1991) and president of the “Informal Group of Developing

Countries” in the GATT (1989-1991):

The present trade regime originated in an Anglo-American conception and was

part of a major restructuring of the international order after the Second World

War, together with the institutions created at Bretton Woods—the international

Monetary Fund and the World Bank—and, in its political dimension, the United

Nations. This structure, particularly its socio-economic aspect, was able to be

built only because there was a hegemonic economic power—the United States of

America, which was determined to be the foundation of the institutional expres-

sion of beliefs regarding free trade that were deep-seated in the inherited values

of the ruling Anglophile and internationalist elite of that country’s East coast

(Ricupero, 1998).

3

Few years before the beginning of Uruguay Round, we can see many changes in the

world trade negotiations. During an US temporary relative decline in 1980s, the Reagan

administration (1981-1989) changed its trade policies regarding US traditional trade part-

ners (Evans, 1989). Domestic protectionism was condemned based on the idea of unfair

trade. US trade partners were pushed to liberalize their trade. At the same time, the US

started to advocate deep reforms in the international trade system. The Uruguay Round

(1986-1994) was, in large part, pushed forward by the US international pressures (Bhagwati

and Patrick, 1990; Bhagwati, 1990) and the WTO is the most important result of the trade

talks.

International trade changes rarely happen without affected countries defending them-

selves. Many times, they form coalitions. For this reason, coalitions increasingly became

active and visible in the Uruguay Round. During the negotiations, most coalitions were

nicknamed bloc-type coalitions because they wanted to blockade many proposals pushed

forward by the US and other developed nations (Narlikar, 2003). One of the most visible

developing country coalitions was the G-10, in which Brazil and India played a prominent

role. The G-10 was composed of countries with diverse interests that made impossible to

this group keeps its unity until the end of the Uruguay Round negotiations (Abreu, 1994).

The pre-launch phase of the Uruguay Round also saw the formation of other coalitions of

developing countries, such as Cafe au Lait and the G-20 (Narlikar, 2003, p. 39). Developed

nations formed their own coalition—called Quad (Canada, European Community, Japan,

and the United States)—to push their liberalization agenda in the Uruguay Round, while

being reluctant to push the opening of agricultural markets in the developed world. During

the Uruguay Round, the European Community (EC) also acted as if it was a coalition,

especially concerning agricultural negotiations and opposing the Cairns Group—formed by

competitive agricultural exporting countries.

Some authors believe that coalitions are “weapons of the weak” because they think

developing countries can counter-balance the power of rich nations only if they increase

their bargaining power through coalitions (Drahos, 2003; Braithwaite, 2004). During the

Uruguay Round developing countries that were leading coalitions—for example, Brazil and

India when they formed the G-10—were not weak countries since they already had a com-

paratively large real GDP. Even back them, countries such as Brazil and India were widely

viewed as leaders of the developing world and important trade nations (Preeg, 1995; Hurrell,

2006; Burges, 2013).

4

Nowadays Brazil and India are called emerging powers or would-be great powers (Schirm,

2010; Hurrell, 2006). During the Uruguay Round, they were weaker in comparison with

great powers of Europe and the United States than now. But still stronger than most

countries in the world. They were, however, unable to block advances in areas such as ser-

vices, investments and property rights—that resulted in major agreements within the WTO

system. They were not able to include the issue of agriculture subsidies in the Uruguay

Round as well. The Cairns Group put the defense of agricultural liberalization on the

Uruguay Round agenda, with an intermediate position between the proposals of the United

States and the European Community (EC) (Cooper, Higgott and Nossal, 1993; Higgott

and Cooper, 1990). After many years of negotiations, however, the US and the European

Community agreed in postpone the liberalization of agricultural products and the Cairns

Group has reduced its own bargaining power. The agreement was celebrate in November

1992 and is known as Blair House Agreement(Meunier, 2000; Veiga, 2005).

After the failure of the Seattle Ministerial Conference in 1999, there was an increased

demand to change the WTO. Public riots and demonstrations outside the deliberations

raised concerns about the internal legitimacy of the WTO. Representatives from the de-

veloping world complained about the marginalization from key decision-making processes

(Schott, 2000; Levi and Murphy, 2006). With a similar rationale to that of the Uruguay

Round, developing countries recognized that they could achieve more by joining economic

forces through coalitions (Narlikar, Daunton and Stern, 2012, p. 172).

During the Doha Round, developing country coalitions were successful in proposing and

defending their narrow or issue-based agenda (Narlikar, 2003). Starting with the Doha Dec-

laration in 2001, they were successful in stating that intellectual pharmaceutical property

rights were not more important than the lives of individuals around the world, especially

the ones living with HIV/Aids and other infectious diseases in developing countries (Sell

and Prakash, 2004; Odell, 2009). Again, developing countries that were advocating this

agenda were not the weakest economic countries in the world. They were at minimum

regional powers. Brazil, India and South Africa were singled out as the main leaders and

winners of the Doha Declaration (Sell and Prakash, 2004; Odell, 2009).

During the Cancun Ministerial Conference in 2003, the G-20 maintained a unified po-

sition around three sets of demands: 1) greater market access in the North; 2) reduction of

developed countries exports subsidies and domestic support mechanism; and 3) a defensive

position to protect agricultural markets of developing countries (Narlikar, Daunton and

5

Stern, 2012, p. 175). The G-20 was also successful in obtaining the support of other smaller

coalitions, such as the G-4, the Five Interested Parties, the G-6, and the G-7. Once more,

the leaders of the G-20 were strong developing nations, such as Brazil, India and China

(Narlikar and Tussie, 2004).

There are many other coalitions working in the WTO. As we can see in Figure 1, all

countries that were at the WTO until 2008 joined at least once a coalition. Many coalitions

are less prominent than the ones we mentioned here (see Table 1). Countries that join

coalitions more are neither weak nor middle powers. Coalitions are, then, another tool

mainly in the hands of powerful countries.

III. Cooperation at the GATT/WTO system

Starting with Olson (1965), collective action literature argues that there is a tendency

for systematic exploitation of the great by the small for the payment of collective benefits.

Olson and Zeckhauser (1966) illustrates the Olson’s theory of collective action with exam-

ples of the NATO and the United Nations. According to them, the United States bore an

excessive cost to promote the collective security of the NATO states. Moreover, smaller

NATO members dedicated small percentages of their budgets for this common goal. Simi-

larly, they argue that the UN urges major economies to bear a disproportionate portion of

the costs of maintaining the institution when compared to least developed countries.

The debate on whether larger groups can be successful working together and promoting

their interests is the trademark of collective action studies for at least four decades (Olson,

1965; Chamberlin, 1974; Hardin, 1982; Ostrom, 1990; Esteban, 2001). Only implicitly, the

mainstream of the collective action literature suggests that smaller members join groups

or coalitions more often because they benefit from public goods paying a small part of the

production costs. In this regard, our findings contribute to the collective action debate by

empirically showing that at the international trade system stronger countries have incentives

to join coalitions.

Within a different framework, Keohane (1969, p. 295) proposes a typology based on

the role that each type of country has on the international system, dividing all countries

into four categories. The first type is the “imperial power” in a unipolar system or the

two major powers in a bipolar system. They are large states that have a central role

in shaping the international system by themselves. Second, there are states that do not

dominate the system individually, but are able to significantly influence it through unilateral

6

or multilateral action. Third, middle powers are not able to affect the system in isolation,

but may be relevant when they act collectively. At last, there are small states that cannot

influence the international system, except when in large groups. Independently, each small

state has a minimal influence within the system.

After Keohane’s typology, a literature has emerged suggesting that middle powers have a

more collectivist behavior than other groups of countries. It is assumed that middle powers

exert a relevant influence in the international arena only through collective action (Higgott

and Cooper, 1990; Cooper, Higgott and Nossal, 1993). Consequently, middle powers would

join coalitions and international institutions more often than either weak and powerful

countries. While weak countries are largely irrelevant, powerful countries are capable of

acting unilaterally without having to make concessions to other coalition members.

Finally, many studies predict that democratic regimes are more likely to engage in free

trade agreements (Mansfield, Milner and Rosendorff, 2000; Dai, 2002; Mansfield, Milner

and Rosendorff, 2002; Mansfield and Reinhardt, 2008). A common argument is that demo-

cratic countries seek to protect wealth gained through international trade because they can

specialize in goods they have a comparative advantage, whereas they import from coun-

tries that have a comparative advantage they do not have (Polachek, 1997; Edward, Milner

and Rosendorff, 2002a; Baier and Bergstrand, 2007). To increase the probability of re-

maining in office, democratic leaders also use international cooperation to signal to voters

their willingness to implement pro-trade welfare-enhancing policies (Edward, Milner and

Rosendorff, 2002b; Mansfield, Milner and Pevehouse, 2008; Mansfield and Milner, 2010).

At last, democracies are also more able to establish credible commitments, which makes

international cooperation more likely (Leeds, 1999; McGillivray and Smith, 2008; Martin,

2000).

As for the main question of this paper, we find that economically stronger countries

join trade coalitions more often than weak ones. We interpret our main finding importing

explanations from economic literature on “transaction costs” (Coase, 1960; Williamson,

1979; North, 1990). Together with the transaction costs, we also take into consideration the

debate about uncertain and asymmetrical gains from participating in GATT/WTO system

(Rose, 2004; Subramanian and Wei, 2007; Tomz, Goldstein and Rivers, 2007). Moreover,

normative changes and law-making can also be considered gains from participating in the

GATT/WTO (Finnemore and Toope, 2001; Goldstein et al., 2000). The literatures on

credible commitments (Leeds, 1999; Evans, Jacobson and Putnam, 1993) and compliance

7

(Downs, Rocke and Barsoom, 1996; Simmons, 1998) are also full of examples about the

difficulties to enforce agreements between countries. Costs are even higher when agreements

are informal—as is the case of GATT/WTO bargaining coalitions (Leeds, 1999; Evans,

Jacobson and Putnam, 1993; Downs, Rocke and Barsoom, 1996; Simmons, 1998).

Transaction costs can be divided into three categories: 1) search and information costs,

2) bargaining and decision costs, and 3) policing and enforcement costs. Search and in-

formation costs are costs to locate the other party one wants to deal with and obtain the

information necessary to negotiate. Bargaining costs are the costs to come to an acceptable

agreement with the other party about the division of surplus, drawing up an appropriate

contract. Policing and enforcement costs are the costs of making sure the other party sticks

to the terms of the contract. If it is necessary, one party can take a legal action to enforce

the contract (Coase, 1960; Dahlman, 1979). Dahlman (1979, p. 148) argues that all three

costs can be reduced to a single one—“resources lost due to lack of information.” Similar to

other studies, we extend “transaction cost” arguments to international relations (Keohane,

1989; Gilligan, 2010; Baccini, 2012).

Larger economies join coalitions more because they are better equipped to pay the

“transaction costs” involved in bargaining together in the GATT/WTO. The “transaction

costs” described are well-documented in the WTO literature. Tussie (2009) edited a whole

volume about the importance of research in trade negotiations. She argues that developing

countries have less resources to gather and analyze information. Thus, coalitions work as

a division of labor, where stronger members provide information for their trade partners

in exchange for their support. Baccini (2012) argues that we can measure bargaining costs

through the length of trade negotiations. The number of countries involved in a negotiation

is also a source of “transactional costs” (Olson, 1965; Olson and Zeckhauser, 1966) as more

countries with different trade preferences need to be satisfied with the final agreement.

The GATT/WTO system has increased the number of countries involved in negotiations.

Different countries have different interests and, as a result, negotiations have become more

and more complex at each round (Barton et al., 2008; Hoekman and Kostecki, 2010).5 The

time span of the trade rounds have increased proportionally.6

5The single undertaking is a foundational rule of the GATT/WTO system. Virtually every item of thenegotiation table is part of a whole and indivisible agreement and cannot be realized separately. As citedat the WTO website: “nothing is agreed until everything is agreed.”

6The first trade round in Geneva in 1947 took 7 months with 23 countries discussing only tariffs’ reduction.The Annecy round in 1949 took 5 months to finish, where 13 countries had also discussed tariffs. The thirdround (Torquay - 1950) took 8 months and 38 countries debated the same topic. At the fourth trade roundin 1956 (again in Geneva), it was needed 5 months to complete an agreement; and 26 countries discussed

8

Our causal mechanism is, then, summarized as the following. Large economies benefit

more from the international trade system and can pay the cost to make sure they continue

to benefit from it either by obtaining material gains or by shaping international rules and

practices. Collectively negotiating is not only costly, but also risky. At the onset, nobody

knows the results of trade rounds and the unintended consequences of trade agreements.

IV. Data and Variables

Coalition-formation is a tool to leverage countries’ bargaining power in the GATT/WTO

negotiations. Countries join a coalition when they share a common identity or interest. Our

dependent variable—coalition entry—was collected from the “Groups in the Negotiations”

section of the WTO website and through the reading of qualitative case studies (see Intro-

duction). It measures the number of times each country joined a trade coalition each year.

We have data on coalitions from 1982 to 2008. We observed our last coalition in 2008, when

the Doha Round was in the middle of a deadlock because of disagreements between some of

its main players on agricultural issues—i.e., the United States, the European Union, India

and Brazil. Consequently, few coalitions have been active until the time of the writing of

this article. We started collecting data in 2008 and we finish it in 2013.

Historical evidences suggest that coalition formation has become more common since

1980s thanks to a series of changes that were taking place on the international scene at

the time: 1) the rise of liberalism of Margaret Thatcher and Ronald Reagan, 2) the de-

cline of American hegemony, 3) the confrontation with newly industrialized countries with

protectionist practices, and 4) the Latin American debt crisis in the 1980s. In a nutshell,

what seemed a relative decline of the US and a change of economic ideology in powerful

countries made the US adopt a more aggressive position towards some of its trade partners

in the developing world (such as, Brazil and India) (Bhagwati and Patrick, 1990; Bhagwati,

1990). As a result, large developing countries started to form more and more coalitions to

tariffs and the admission of Japan. Finally, at 1960, 26 countries discussed tariffs in the Dillon round.The Kennedy round in Geneva increased substantially the number of country participants (62) and thetime to complete the round (37 months) on tariffs and anti-dumping rules. The same happened with theTokyo round, where 102 countries took 74 months to reach an agreement on tariffs, non-tariff measures,and “framework” agreements. During the time of our study (from 1982 to 2008), the trade rounds havebecome extremely long and complex. Transaction costs have been costly for many countries. The UruguayRound, which started at 1986, involved 123 countries that negotiated many trade topics during 87 monthsand fundamentally changed the whole international trade system by creating the WTO in 1994. Finally,the Doha round started in 2001 with 159 countries with many more topics on the table, as for example thestruggle for the liberalization of services and agriculture markets. The Doha talks have only come to anagreement after 12 years of negotiations.

9

defend themselves. All these episodes were important to push forward trade negotiations

that led to the Uruguay Round (1986-1994) (Gilpin and Gilpin, 1987; Barton et al., 2008)

We withdrew from our data set countries that have not participated in the GATT/WTO

system during the time of our study (1982-2008). Countries have become part of our data

set starting at the moment they entered the GATT/WTO system. Coalition activities

are difficult to observe. It is hard to know for sure if a member either stayed or left

a coalition during the negotiations. Historical records are vague. Many times the only

moment when a coalition let the name of its members clear is when it proposes its first

negotiation draft. To supplement the information on the WTO website, the number of

coalition members is measured at the year of the coalition formation based on the reading

of the qualitative literature and of official documents from the GATT/WTO. Unfortunately

we cannot distinguish between short- and long-lived coalitions with our data.

To mitigate problems related to measurement error in our dependent variable, we mea-

sure coalition entry in two different ways: i) all coalitions that have participated at the

GATT/WTO (coaltrade) and ii) coalitions that have just been created for the GATT/WTO

negotiations (coalwto). For “coalwto”, we exclude coalitions that negotiate both inside and

outside the GATT/WTO system; for example, African, Caribbean and Pacific group, Apec

and Caricom, and also regional trade alliances as European Union, Mercosur and Nafta—

see Table 1. Regional integration coalitions were classified as Non-WTO coalitions. Our

analysis with two different dependent variables (coaltrade and coalwto) are not substan-

tively or significantly different. The results were slightly weaker in the second measurement

(coalwto) since we increased the number of zeros in our dependent variable.7

A. Independent Variables

We collected all our independent variables from official databases. The description of

the variables is presented in Table 2. We have two economic variables. Real GDP is our

main explanatory variable. Contrary to the view that suggests that coalitions are weapons

of the weak or of middle powers, we found that the higher the real GDP, the higher is the

probability of countries joining a coalition. We collected data on GDP from Penn World

7We have done the best we can to avoid measurement errors in our dependent variable. However, randommeasurement errors in the dependent variable Y does not bias substantively the estimates of our regressionmodels. It does inhibit our ability to estimate the relationship between Y and our independent variablesprecisely, making the variance larger. Measurement error in the independent variables, on the other hand,always biases the coefficients towards zero—either overestimating negative coefficients or underestimatingpositive ones.

10

Table (PWT 6.3.)

Our other economic variable—trade openness (openk)—was also collected from PWT

6.3 and is measured by the sum of import and export flows divided by GDP of each country

annually. Higher levels of trade openness is also an important predictor of a country joining

a coalition. The result on trade openness is less counter-intuitive because the GATT/WTO

system is ruled by a free-trade principle.8

Because of the debate whether democracy should be a continuous or categorical measure

(Cheibub, Gandhi and Vreeland, 2010; Przeworski et al., 2000; Elkins, 2000; Collier and

Adcock, 1999), we use two measures of democracy. The first is a dichotomous indicator—

democracy versus dictatorship — created by Alvarez et al. (1996) and updated by Cheibub,

Gandhi and Vreeland (2010). The second is a continuous index of democracy—Polity IV

(Marshall and Jaggers, 2002). The Polity IV democracy score varies from -10 (complete

dictatorship) to +10 (full democracy). Table 2 presents descriptive statistics of our main

variables divided by the North-South hemispheres and also for the whole world.

To deal with the long-standing literature on the North and South divide (Raffer and

Singer, 2001), we include a proxy to capture North-South division as a control in our

study. We cannot include the North-South variable in our panel analysis because fixed-

effect models do not allow the inclusion of time-invariant variables. North-South variables

are measured according to the Equator line. Countries below the Equator are classified as

Southern countries. Countries above the Equator are Northern countries. The information

on the Equator line was collected from the CIA Factbook.9 To be clear, the literature on

North-South divide refers to geographic divisions as well as political and economic divisions.

Normally the geographic division is a metaphor for economic and political divisions. For

instance, few authors disagree that New Zealand and Australia are part of the economic

and political bloc of Northern countries; whereas China and India are viewed as Southern

countries. According to the North-South divide literature, usually geographic and political-

economic characteristics overlap. Albeit we are sure that the geographic measure is not

completely fair to such an ample literature, other problems would emerge if we create our

8We also tested if real GDP per capita, size of the government, growth rate of real GDP per capita,battle-related deaths, electric power consumption (kWh per capita), exports (% of GDP), imports (% ofGDP) and total trade (% of GDP) have an effect on the coalition entry. They do not have either asignificant or relevant effect. Conducted at early stages of the research, extreme bounds analysis (Leamer,1985) (a sensitivity analysis technique) and Bayesian additive regression trees (BART) (Chipman, Georgeand McCulloch, 2010) (a machine learning technique) also selected the variables used in this paper as themost relevant.

9https://www.cia.gov/library/publications/the-world-factbook/

11

own measure, since countries move up or down in a scale of power and development (e.g.

Asian Tigers and BRIC nations). As sometimes world changes are not clear-cut, we choose

to use a geographic and more straightforward and time-invariant measure of Northern and

Southern countries.

Finally, the size of members’ delegation in Geneva is used as a proxy for the “transaction

costs” that countries incur when negotiating at the GATT/WTO. Delegates in Geneva is

a mediator variable between real GDP and coalition entry. The WTO used to publish

an annual phone directory for internal use which lists each member’s delegates in Geneva

(Busch, Reinhardt and Shaffer, 2009, p. 562). Unfortunately there was only data available

from 1994−2004—a shorter time-horizon than the one of used in the remaining of our paper

(1982 − 2008). The data used here were collected by Sattler and Bernauer (2011).

V. Methodology

We employ both panel data econometrics and cross-section (pooled) non-linear models

in this study. In panel form, we present results from Ordinary Least Squares (OLS),

Two-Stage Least Squares (2SLS) and both Logistic and Negative Binomial regressions.

To explore our proposed causal mechanism, we present a mediation analysis (Imai et al.,

2011) using multilevel models, where the intercept of our regressions vary yearly (Gelman,

2007). In cross-sectional form, we present graphical results from the non-linear Generalized

Additive Models (GAMs). In this section, we outline some basic features of our models.

A. Panel Data Econometrics

Panel data techniques offer a series of advantages over cross-section analyzes. For exam-

ple, increasing the estimation accuracy and controlling for unobserved heterogeneity (Hsiao,

2003; Cameron and Trivedi, 2005). In this paper, we run OLS, 2SLS, Logistic and Negative

Binomial panel models.

Fixed-effects models include different intercepts for each individual countries. Unob-

served heterogeneity α is no longer a random variable but a parameter to be estimated.

Fixed-effects models allow for unbiased estimation even in the presence of unobserved het-

erogeneity α correlated with the regressors (Wooldridge, 2010, 2012)10

Some econometricians defend the use of OLS models even when we have limited de-

10Random-effects models assume that unobserved heterogeneity α is a random variable distributed re-gardless of the regressors. In our case, Hausman tests show that α is correlated with the regressors and thenonly the fixed effects model is consistent.

12

pendent variables (e.g., Logit) or count dependent variables (Angrist and Pischke, 2008;

Wooldridge, 2012). Our preferred panel model is a simple OLS fixed-effects model.11 The

advantages of OLS models are that they are normally consistent and easier to interpret

than Logit and Negative Binomial models. Below we describe the OLS model:

Yit = αi + βXit + γZit + εit, (1)

where t = 1,..., T and i = 1,..., N . αi is the unobserved time-invariant individual country

effect (country-specific intercept); Yit is the dependent variable “coalition entry” observed

for individual country i at time t; βXit is the time-variant regressor of the log of real GDP;

γZit is a vector of other covariates for individual country i at time t; and εit is the error

term for individual country i at time t.

The Logistic Panel Model is employed when the dependent variable is defined as whether

or not countries enter a coalition each specific year. The relationships between the explana-

tory variables and the probability of entering in coalitions are estimated using maximum

likelihood techniques. Finally, count panel data econometrics is an obvious statistical tool

to tackle the problems posed given the count structure of our dependent variable. The

Negative Binomial distribution is a generalization of the Poisson distribution, allowing a

less restrictive mean-variance relationship (Cameron and Trivedi, 1998). As in our case,

if data is not equidispersed12, Negative Binomial Regression is more suitable because the

ratio between the mean and variance of the dependent variable is also a parameter to be es-

timated. However, the estimation algorithm of Negative Binomial models is less stable and

efficient than other GLM algorithms, such as the Logit models (MacCullagh and Nelder,

1989)). As we show in Basic Results, most results from Logit, Negative Binomial and OLS

models with fixed-effects are similar.

A.1. Instrumental Variables and Identification Strategy

Reverse causality is not a severe problem in our study. As far we know, there is no theory

suggesting that coalition participation has an effect on country’s real GDPs or on any other

main independent variable (political regime and openness to trade). Substantively it is

also not intuitive how participating in a small part of the GATT/WTO activities exert an

impact on a country’s GDP, political regime, openness to trade or any other independent

11With GAM, we easily interpret our models graphically in a cross-section setting.12We conducted a test proposed by Cameron and Trivedi (1998) that shows data is not equidispersed.

13

variable employed by us. The size effect of the relationship between real GDP and coalition

entry is large (see Basic Results). Then, it is unlikely that entering in a coalition can

increase a country GDP in the same proportion.

Country’s real GDP can, however, co-vary with omitted variables that are correlated

with coalition entry, such as the expertise and number of negotiators by each country that

bargains in the GATT/WTO. While an increase in real GDP reduces the incentive to join

coalitions, the increase in organizational capacities become less expensive. Both effects

together cancel out one another. Therefore, our instrumental variable is intended to solve

the problem of omitted variable bias, but it also deals with potential minor problems of

reverse causality.13

Two-stage least-square (2SLS) regressions is a common solution for omitted variable

bias. 2SLS can provide consistent estimate when the instrumental variable is orthogonal to

the dependent variable, once controlled by the effects of endogenous independent variables.

The instrument must be correlated with the endogenous variable and uncorrelated with

the error term. Otherwise, it will suffer from the same problems that we are trying to fix

(Angrist and Pischke, 2008; Kennedy, 2008; Wooldridge, 2010, 2012).

We lagged in seven years the number of natural disasters at country-level and used it

as an instrumental variable for GDP. Lags decrease the probability of correlation between

the error terms of our 2SLS regression (Murray, 2006), which is already unlikely since our

instrument is “as if random.” In other words, “natural disasters” happen by chance and in

an haphazard way (Sovey and Green, 2011).14 There are also not any theory that relates

the number natural disasters with any form of participation in international coalitions. As

the number of natural disasters do not directly influence whether a country enters into trade

coalitions or not, the hypothesis of exclusion restriction is sounding. Countries that had

experienced natural disasters develop better preemptive measures and technologies, what

increases savings and investments to reduce the damages of future disasters that normally

happen infrequently. In turn, investments and savings have a positive effect on GDP in the

long run (Albala-Bertrand, 1993; Skidmore and Toya, 2002; Crespo Cuaresma, Hlouskova

13Previously we adopted a strategy similar to Chen and Nordhaus (2011) and Alesina, Michalopoulos andPapaioannou (2012), who use data on luminosity to correct measurement errors in poor countries’ GDPs.We found a strong correlation between countries’ total number of phone lines and countries’ GDPs. Whenwe include phone-lines as an instrument, the results are similar to OLS regressions.

14Several papers have used natural occurring phenomenon as instruments for political and economic vari-ables(Sovey and Green, 2011). The notorious work of Miguel, Satyanath and Sergenti (2004), for example,use rainfall as an instrument of per capita economic growth to estimate its impact on civil conflict insub-Saharan Africa.

14

and Obersteiner, 2008).15 In sum, any effect of past natural disasters on coalition entry is

expected to happen through GDP.

A.2. Mediation Analysis and Causal Mechanisms

Mediation analysis examine causal mechanisms that underlies a relationship between an

independent variable and a dependent variable through the inclusion of a third mediation

variable. Rather than suppose a direct causal relationship between the independent variable

and the dependent variable, a causal mediation model shows that the independent variable

influences a dependent variable indirectly through a mediator. The indirect effect represents

the expected causal mechanism and the direct effect represents all the other mechanisms.

The total effect is the sum of the average causal effect and the direct causal effect (Imai

et al., 2011).

We argue that countries with higher aggregated economic power have a higher likelihood

of joining bargaining coalitions in the GATT/WTO. Our causal mechanism is based on the

idea of “transaction costs.” Large economies can pay for the uncertain and uneven results

of the GATT/WTO negotiations, since they have enough resources to gather information

on trade, to hire law firms and skillful diplomats and negotiators.

The idea of transaction costs is difficult to measure accurately. Following Sattler and

Bernauer (2011), Bown (2005) and Guzman and Simmons (2005), we use the number of

delegates in Geneva as a proxy of bureaucratic capacity of countries bargaining at the

GATT/WTO. As the number of delegates only captures some of transaction costs involved

in negotiating together in the GATT/WTO, we are probably measuring our mediator with

errors and underestimating its true indirect effect. For example, many delegates work both

at the GATT/WTO and at other international institutions—notably the United Nations.

Some countries might have fewer delegates in Geneva, but hire more independent nego-

tiators, and law and consulting firms. At last, number does not mean quality—a higher

number of delegates do not translate directly into a more efficient gathering of information

and negotiating skills (Busch, Reinhardt and Shaffer, 2009). Despite potential problems of

using the number of delegates in Geneva as a proxy for transactional costs, it is unlikely

that it systematically underrepresents other types of transaction costs. Probably, it hap-

pens in the other way around. Countries that can pay for a large diplomatic mission in

15Not shown in this paper, scatterplots with loess lines and several GAM specifications suggest that therelationship between seven years lag of natural disasters and real GDP is monotonic, which makes theassumption of linearity reasonable.

15

Geneva are also more able to pay for other types of transaction costs. As we said in the

Data and Variables section, the WTO only provides the number of delegates in Geneva

from 1994 to 2004—a shorter time horizon than the period of our research, which weakens

our statistical results.

There is a dense debate in the statistical literature on mediation analysis (Baron and

Kenny, 1986; Sobel, 1982). In all mediation methods the researcher must make strong

assumptions, especially when using observational data. Nonetheless, investigating causal

mechanisms is too important to be left out. The statistical debate on different mediation

analysis methods is beyond the scope of this work. Here, we employ the method proposed

by Imai et al. (2011), which involves two stages. We first use a linear multilevel regression

to estimate the effect of log of GDP on the number of delegates.16 Second, we use a Logit

multinomial regression to estimate the effect of log of GDP on coalition entry, controlling for

democracy, trade openness and North-South hemispheres. In the first equation we assume

that the number of delegates is random if countries have similar GDPs. In the second

equation we assume that countries join coalitions randomly if they have similar GDPs.

The mediation analysis relies on the assumption that there is no unmeasured mediator

influencing both coalition entry and the number of delegates in Geneva. In the Basic

Results section, we further detail our findings.

B. Cross-Section Analysis

B.1. Generalized Additive Models (GAMs)

In this paper, not only we want to verify a relationship between coalition entry and

our independent variables, but also to visualize its functional form. One advantage of

Generalized Additive Models (GAMs) is to display non-linear marginal effects graphically.

Unlike linear models that assume a global linear relationship between the dependent and

independent variables, GAM allows the relationship varies locally over the range of the

dependent and independent variables (Hastie and Tibshirani, 1990; Beck and Jackman,

1998). We estimate GAM using a Bernoulli distribution with a Logit link function. The

GAM models establish a smoothing function of the relationship between the dependent and

independent variables, given the control variables. Below, we formally present our main

non-linear model:

16Gelman (2007) is the best reference we know on multilevel models for political scientists.

16

log(πi

1 − πi) = α+

p∑j=1

fj(Xj) + εi, (2)

where πi is the probability that the dependent variable “coalition entry” at individual

country i is 1 rather than 0; α is a constant; p is the number of independent variables; εi

is the error term which is independent of the Xj , var(ε) = s2, E(ε) = 0 and i = 1,..., N .

fj(Xj) is a smooth function. fj replaces the linear function βXit that we see in our linear

model (see Equation 1).

GAM is a semi-parametric model and is more general than the GLM. It is parametric

when we assume a distribution for the dependent variable, modeling its average. It is non-

parametric when the average is modeled through smoothing functions, which are functions

that estimate the functional form of the relationship between the independent variable and

the dependent variable. As we can include linear and categorical variables in GAM, below

we extend the Equation 2:

log(πi

1 − πi) = α+

p∑j=1

fj(Xj) +

q∑j=1

βkZij , (3)

where q is the number of non-smoothed terms to be included. The most simple smooth

function is a moving average. It averages neighborhoods of points around the target value

to produce an estimate. More sophisticated functions employ weighting to points as they

move away from the target value. Loess, for instance, uses locally linear fits and is explicitly

local in nature, the degree of locality depending on the span of data points used.

Here, we use cubic smoothing splines, but loess provides similar results. Regression

splines “offer a compromise by representing the fit as a piecewise polynomial” (Hastie and

Tibshirani, 1990, p. 22). Polynomials are added together by knots, and the polynomial

curves between regions join smoothly at each region break point. Cubic smoothing splines

reduce the compromise between goodness-of-fit and degree of smoothness. If the number

of degrees of freedom is low, the spline is smooth. If it is high, a less smooth spline are gen-

erated. Unlike Beck and Jackman (1998, p. 610), we have not constrained our smoothing

parameters, relying on the mgcv package from R to automatically choose smoothing param-

eters so as to minimize prediction errors. We have already presented other linear, binary

and count dependent variable models, so that here we truly want to “get the mean right.”

With automatically chosen smoothing parameters, finding that our main variable—real

17

GDP—has an almost global linear relationship with coalition entry makes our results less

controversial (or manipulative) from the statistical point of view. Finally, we use a Logit

transformation function to visualize the relationship between our dependent and indepen-

dent variables in a probability form. The use of probabilities facilitates the interpretation

of our results.

VI. Basic Results

A. Panel Data Econometrics

The log of real GDP increases the likelihood of a country joining a coalition at the

GATT/WTO across the two different ways we measure our dependent variable. Table 3

presents linear models in which “coalition entry” includes both all coalitions that participate

in the GATT/WTO (Panel A) and the ones that have only been created for negotiating at

the GATT/WTO (Panel B). F-statistics corresponds to the test of the null hypothesis that

the coefficient of the excluded instrument equals 0. Seven years lag of natural disasters is

a strong instrument as F-statistics varies from 24.85 to 12.65. Staiger and Stock (1997)

suggest F ≥ 10. The Hausman test of endogeneity compares the OLS and the 2SLS panel

data estimators. We rejected the null hypothesis that both estimators are consistent, so

that the 2SLS regression is more appropriated. All specifications include country fixed

effects and bootstrap standard errors with 400 resamples.

Columns 1 through 3 present model specifications that estimate the impact of real

GDP on coalition entry. In column 1 (Panel A) we regress the seven-years lagged number

of natural disasters on the potential endogenous variable real GDP (First Stage of the 2SLS

regression). Column 2 (Panel A) shows the effects of real GDP on coalition entry without

the use of an instrument. The probability of a country joining a coalition increases by 63%

when real GDP increases by one standard deviation. After being instrumented, the effect of

1 standard deviation increase in GDP boosts the probability of a country joining a coalition

by 142%. The Hausman test (476.00; Prob > χ2 = 0) suggests that both the OLS and the

2SLS models are dissimilar. Then, when we are not including an instrumental variable we

are being conservative about the true effect of the real GDP on coalition entry.

In column 2 (Panel B) we find that one standard deviation increase in GDP increases the

probability of a country joining a coalition by 61%. When we use the instrument (column

3), 1 standard deviation increase of GDP boosts the probability of a country joining a

coalition by 139%.

18

In columns 4 through 6 we run models specifications that estimate the impact of GDP

on both measures of coalition entry (Panel A: coaltrade and Panel B: coalwto), controlling

by the dichotomous indicator of democracy. Again, there is substantive difference among

the models with and without the natural disaster instrument. In column 5 (Panel A) we

present the model without the instrument. Once again, the OLS model is a much more

conservative estimation of the true effects of GDP. The political regime indicator appears

to have an insignificant impact and do not affect the GDP coefficient. In column 5 and 6

(Panel B) we find the same results: democracy is not an important determinant of coalition

entry at the GATT/WTO and it does not affect the GDP coefficient.

Columns 7 through 9 (Panel A) show similar results, but now we use a continuous index

of democracy (Polity IV) as a control variable. The effect of GDP remains similar and the

democracy index has an insignificant effect on coalition entry. Once again, the OLS model

underestimates the true effects of GDP.

Finally, columns 10 through 12 includes the trade openness as a control variable.

Columns 11 and 12 show that GDP remains as a good predictor of coalition entry. At

Panel A, an increase of one standard deviation on GDP increases the probability of a coun-

try joining a coalition by 50%, while at Panel B its impact is 48%. Apart from that, on

both Panels A and B trade openness has also a significant impact on its own, reducing the

effect of GDP. Columns 11 and 12 shows that an increase of one standard deviation on the

level of trade openness increases the probability of coalition entry by 31% on Panel A and

33% on Panel B. The results for trade openness was expected because the GATT/WTO is

a pro-open trade institution.

Results are stable across all our model specifications: all are statistically significant,

most of them at 99% confidence interval. The effect size of GDP and trade openness are

large. As expected, Panel B present weak results because we have less observations when

we exclude coalitions that negotiate at the GATT/WTO but were not create with this goal.

Table 4 repeats the specifications shown in Table 3 with both measures of coalition

entry—“coaltrade” in Panel A and “coalwto” in Panel B. In both panels we use Logit and

Negative Binomial models. Since our instrument provides even stronger results than the

ones from other models, we choose not to include instrumental variables in more complex

models. Therefore, we can be more conservative in our estimations and to make inferences

for all set of countries independent of their numbers of natural disasters. One disadvantages

of instrumental variables is that they only estimates local average treatment effects (LATE)

19

(Angrist and Pischke, 2008). In our case, we estimate that countries that have an increase

in their GDPs because of previous cases of natural disasters join coalitions more often. The

results from Table 4 are consistent with linear models. The impact of the log of GDP is

statistically significant in 15 of 18 model specifications. The exceptions are when we use

Negative Binomial regressions in Panel B (dependent variable “coalwto”).

Table 5 presents OLS, Logit and Negative Binomial models with fixed effects and with all

independent variables used in Tables 3 and 4. All models include GDP, either a dichotomous

indicator (democracy versus dictatorship) or a continuous index of political regime (Polity

IV), and trade openness. GDP is the only variable that has remained statistical significant

in all model specifications, making clear that it is the most important independent variable

in our study. Out of four Logit models, trade openness only barely achieves statistical

significance at 10% in one of them—Panel B; column 3.

In sum, our results show that GDP is the most important variable in our study. It

presents a large size effect and it normally reaches statistical significance at 0.1% level.

Trade openness is also an important predictor of coalition entry. In general, however, its

size effects are smaller than the real GDP and it is less often statistical significant. Finally,

both measures of democracy are weak predictors of a country joining a coalition. They are

rarely statistical significant and their effect sizes are small. Nevertheless, the positive sign

of the democracy coefficients is consistent with our theoretical expectations.

B. Mediation Analysis

The mediation analysis conducted here is divided into two steps. We use linear multilevel

regression to estimate the effect of log of GDP on the number of delegates in Geneva.

Second, we use a Logit multinomial regression to estimate the effect of log of GDP on

coalition entry, controlling for democracy, trade openness and North-South hemispheres.

In the two models we created one intercept for each year from 1994 to 2004. We use a

time-horizon shorter than the one used in the remaining of the paper (1982—2008) because

there is no data available on the number of delegates for a longer period. In the first

equation, we assume that the number of delegates is random if countries have similar

real GDPs, controlling for our covariates. In the second, we assume that the number of

times countries join a coalition is random if they have the same real GDP, also controlling

for the same covariates. The last assumption—called sequential ignorability—states that

there is no observed or unobserved (mediation) covariates that affects both coalition entry

20

(“coaltrade” and “coalwto”) and the number of delegates in Geneva. We use the mediation

package from R to estimate our models. We use bootstrap standard errors with 1,000

resamples in all models.

When the outcome model is a GLM model, such as our Logit multinomial regression,

the ACME and direct effect (ADE) estimates will differ between the treatment and control

conditions (Tingley et al., 2014, p. 8). Because point estimates of the control and treatment

groups largely overlap for both ACME and ADE, we only interpret the averages of ACME

and ADE. If we add up ACME to ADE, we obtain the Total Effect—see Figures 2f, 3f, 4f and

5f. Dots represent point estimates. Horizontal lines depict 95% confidence intervals—with

a vertical line at zero to facilitate interpretation.

We have four different model specifications. In Figure 2, the dependent variable—

coalition entry—includes all countries that participates in coalition at the GATT/WTO

(coaltrade) and we use a dichotomous measure of democracy as a control variable. In Figure

3, the dependent variable only includes countries that participate in the GATT/WTO

(coalwto) and we use the same measure of democracy. In Figure 4, we use the dependent

variable “coaltrade” and a continuous measure of democracy (Polity IV). Finally, in Figure

5 we use the dependent variable “coalwto” and also a continuous measure of democracy.

Our first mediation model at Figure 2 has 1451 observations. The probability of the

ACME is 1.33%—statistical significant at zero [95% CI: 0.00384, 0.02466]. The direct

effect (ADE) of log of real GDP decreases the probability that a country joins a coalition

by −2.2%—also significant at zero [95% CI: -0.03583, -0.00745]. Total effect decreases

the probability by −0.84% [p-value: 0.1; CI: -0.01651, 0.00128]. Our proposed mechanism

(ACME) correspond to a −141.6% change of the total effect—[p-value: 0.1; CI: -10.97412,

13.56305]. In Figure 3, we also have 1451 observations. ACME has a probability of 0.7%

[p-value: 0.05; 95% CI: 1.75e-05, 1.74e-02], whereas the decreasing probability of ADE is

−0.5% [p-value: 0.5; 95% CI: -0.0211, 0.00365]. Total effects add up to 1.9% [p-value: 0.4;

CI: -0.00735, 0.00689]. Our mechanism corresponds to 119% of the total effect [p-value: 0.4;

CI: -23.1, 19.9]. In both Figures 2 and 3, we employ a dichotomous measure of democracy.

The following two mediation models have 1253 observations—see Figures 4 and 5. In

Figure 3b, the probability of the ACME is 1.45%—statistical significant at zero [95%

CI: 0.00256, 0.0291]. ADE has a marginal effect of −1.6%—[p-value: 0.09; CI: -0.0368,

0.000689]. The marginal effect of the total effect is −0.12% [p-value: 1; CI : -0.0127,

0.00623]. The proportion mediated by the number of delegates is 32.5% [p-value: 1; CI:

21

-44.2, 32.8]. In Figure 5, ACME has a probability of 1.5%—significant at 1% [95% CI:

0.00275, 0.0288], whereas the marginal effect of ADE is −1.62% [p-value: 0.06; 95% CI:

-0.0354, 0.000316]. The marginal total effect is −0.154% [p-value: 1, -0.0133, 0.00607]. Our

causal mechanism changes the total effect by −61% [p-value: 1; CI: -34.2, 41.2]. As we

employ a continuous measure of democracy (Polity IV) in both Figures 4 and 5, we reduce

even more our data set in comparison with the first two models that employed a dichoto-

mous measure of democracy—see Figures 2 and 3. As a result, there are more uncertainty

about the estimates in the models that employ a continuous measure of democracy as a

control.

Having an intercept for each year from 1994 to 2004 does not meaningfully alter our

results. As expected, the uncertainty of our estimates—measured by confidence intervals—

are even larger when yearly sub-setting the data (see Figures 2b, 2c, 3b, 3c, 4b, 4c, 5b,

5c; Figures 2d, 2e, 3d, 3e, 4d, 4e, 5d, 5e; and Figures 2f, 3f, 4f; and 5f). Nevertheless, the

direction of our coefficients, showing that our results are consistent throughout the years.

In sum, the effect of log of GDP on coalition entry mediated by its number of delegates

in Geneva (ACME) presents a probability that goes from 1.5% to 0.7%—always highly

statistical significant. The average direct effect of log of GDP on coalition entry goes

from −2.2% to −0.5%—presenting more uncertainty in its estimates. The results suggest

that the effects of real GDP on coalition entry is mediated by the number of delegates in

Geneva. Since the ACME is always positive and the ADE is negative, the total effect is

close to zero, but our expected mechanism is responsible for around 100% change of the

total effect. Our small, but statistical significant average mediation causal effects (ACME)

pale in comparison with other results presented by us. We should evaluate them cautiously,

though. Because of a small sample size, some of our inferences are uncertain. Moreover,

as we argue in the Mediation Analysis and Causal Mechanisms, the number of delegates

in Geneva is only a proxy of some “transaction costs” a country pays when negotiating

collectively at the WTO. Our results are robust to other model-based mediation analysis

and other model specifications not shown in this paper.17 Also not shown in this paper,

both real GDP and the number of delegates have a significant effect on coalition entry when

using panel fixed-effects models. Nonetheless, the effect of both variables decrease when

they are put together in the equation.

17Sensitivity analysis is not available for multilevel models in mediation package. In other models, acorrelation at around 0.1 between the errors of the first and the second regression equations changes thesign of our ACME coefficients from positive to negative.

22

C. Non-Linear Models with GAM

GAM allows estimation of smoothing functions rather than linear coefficients for each

independent variable. The marginal effect of each variable is a smoothing function estimated

by GAM, keeping other independent variables fixed. Our fitted GAM models are analogous

to a usual logistic regression. The dependent variable is an indicator of whether a country i

joined a coalition or not. Logit was used as a link function of the average of the dependent

variable with the independent variables. Since we have a Bernoulli distribution [0, 1], we

choose a Logit link function for simplicity because other link functions do not change our

results significantly—other obvious choices are Probit, Tobit or Cloglog distributions. As we

analyze GAM models graphically, a Normal distribution does not improve the interpretation

of our results. A Negative Binomial distribution, on other hand, it is more difficult to be

fitted in the GAM framework and do not change our results significantly. Here we use all

independent variables from the Panel Econometrics section. We also include a dichotomous

control variable to divide our data set according to the North-South hemispheres—see

Figures 6 and 7. We first present GAM results from Figures 6a to 6c—which includes

GAM estimates for all coalitions that have participated at the GATT/WTO negotiations

(coaltrade). Second, we analyze the results from Figures 7a to 7c that only take into account

coalitions that were built inside the GATT/WTO (coalwto). In all our models, we changed

the y-axis scale to display probabilities.

Figures 6a, 7b and 6c present—respectively—the non-linear marginal effects for real

GDP, trade openness and Polity IV on “coalition entry” (coaltrade).18 Vertical axis scales

were transformed by a Logit link function into probability intervals [0, 1]. Each graph shows

the estimated value of the degrees of freedom of each smoothing function. In GAM, degrees

of freedom indicate the complexity of the estimated smoothing function. When the degree

of freedom of the marginal effect of a variable is equal to 1, the relationship is linear and

the marginal effect is equivalent to a linear predictive model without a smoothing function

(see Figures 7a and 7b).

Figure 6a shows that the marginal effect of real GDP increases monotonically. Keeping

all other variables constant, the higher the real GDP, the more likely a country will join

a coalition. The shape of our smoothing functions is robust to different model specifica-

tions. Furthermore, the relationship is close to linear, which suggests that a global linear

18One limitation of the GAM framework is that it does not graphically display dichotomous independentvariables, such as the case of our dichotomous measure of democracy.

23

function—instead of a smoothing function—is the most appropriated method. Assuming

linearity is theoretically more parsimonious. The assumption of linearity works quite well

for our the relationship between real GDP and coalition entry in the two ways we measure

our dependent variable. The probability of the lowest real GDP countries in our data set

to join a trade coalition is 20%. Middle income countries (or middle powers) present a

probability of less than 60%. The wealthiest countries join coalitions with more than 80%

probability. In sum, richest countries in aggregated terms are the ones that have partici-

pated more in coalitions within the GATT/WTO system. Middle—and especially smaller

powers—are being left behind.

Figure 6b display marginal effects of trade openness. The relationship is positive. Coun-

tries more open to trade are also more likely to join a coalition—controlling for other co-

variates. There is a curvilinear effect when the log of trade openness is above 2. The

effect correspond to the highly open economy of Singapore. When the outlier Singapore

is dropped from the sample, the effect of the log of trade openness becomes almost linear.

Our shaded confidence intervals also becomes larger after the 2.5 threshold.

In Figure 6c—controlling for our other independent variables—the effect of Polity IV

index is clearly not linear. It presents an almost inverted U-shaped curve such as Figure

6b. Apparently, there is a threshold value close to −5. At countries with low levels of

democracy, more democracy means more cooperation. The effect is positive and locally

linear for countries with lower levels of democracy. As for countries above the threshold,

the effect of more democracy is virtually nonexistent. Actually, there is a slightly downward

trend for highly democratic countries.

Figures 7a, 7b, and 7c present GAM estimates for our measure of coalition entry

that considers only coalitions that were built inside and for the GATT/WTO negotia-

tions (coalwto). The results only change slightly. The effects of real GDP is still almost

(globally) linear (Figure 7a). At the lower bound of the distribution, the probability of

a country joining a coalition is 20%. At the upper bound the probability is over than

80%. Trade openness also presents an almost inverted U-shaped curve (Figure 7b). The

probability of a closed economy to join a coalition is around 40%. Finally, again, Polity

IV presents a local linear effect on coalition entry for dictatorships open their political

regimes—the probability of joining a coalition goes from 20% to 60% between −10 and

−5 in the Polity IV score. After that, there is an incremental downward trend. In the

end of the distribution, the trend slowly accelerates—see Figure 7c. There is a curvilinear

24

threshold at around −5 level of democracy in the upper bound of the distribution. After

the 5.5 threshold, there is a strong downward trend, where lies the countries with higher

levels of democracy. The results suggest that only dictatorships that increasingly become

more democratic cooperate more in the GATT/WTO system.

VII. Conclusion

We found strong and robust evidences that larger economies tend to join coalitions more

often than smaller ones. There is a linear and strong relationship between countries’ higher

real GDP and coalition participation within the GATT/WTO system. The magnitude of

the size effects is around 60%. We have obtained similar results with different statistical

methods, several model specifications and two different ways of measuring our dependent

variable. In addition, when we use non-linear regressions, the relationship between GDP

and coalition entry is kept almost linear.

Our main finding, then, challenges the “weapons of the weak” argument (Drahos, 2003;

Braithwaite, 2004). Coalitions are not one of the few weapons in the hands of weaker

countries. They are another tool that is already more employed by strong economies—

either from the developing or from the developed world. Since the relationship between

real GDP and coalition entry is linear, middle economic powers also do not present a

distinctive collectivist behavior if we compare them to large economies (Cooper, Higgott

and Nossal, 1993; Higgott and Cooper, 1990).

Countries more open to trade also participate in coalitions often in contrast with closed

economies because the GATT/WTO is an institution that stimulate free-trade between

nations. More unexpected is finding few evidences that democracies cooperate more than

undemocratic countries across two different measures of democracy and several statistical

methods and model specifications. Granted, we only selected countries that are members

of the GATT/WTO and we cannot generalize our findings on political regime to other

domains. As a caveat, the positive sign of the panel regression coefficients and a local

linearity in GAM models do not discard the “democratic peace” literature all together

(Edward, Milner and Rosendorff, 2002a; Baier and Bergstrand, 2007). What the nonlinear

findings show is that the “democratic peace theory” is more complex than linear (or GLM)

models assume. Actually, Beck and Jackman (1998) proposed non-linear models to refine

the “democratic peace theory” in one of the first and few articles in political science to

25

employ the GAM methodology.19

The “transaction costs” literature (Coase, 1960; North, 1990; Williamson, 1979) help us

to make sense of our main finding. Bigger economies tend to participate more in coalitions

because they have enough resources to pay for the high “transaction costs” of collectively

participate in complex and long negotiations. Among other costs, countries have to keep

diplomatic missions in Geneva. Mediation analysis suggests that our causal mechanism

is plausible. Bargaining in the GATT/WTO is also a risky business. As coalitions are

groups competing in negotiations with other countries and coalitions (Esteban, 2001), vic-

tory is not guaranteed. Economic gains from participating in the GATT/WTO is also

debatable (Rose, 2004; Tomz, Goldstein and Rivers, 2007) and probably uneven in favor

of rich-developed countries (Subramanian and Wei, 2007), discouraging weaker economies

to collectively bargaining in the GATT/WTO negotiations. We, finally, urge researchers

to further test our causal mechanism with other research designs and methods. Other

measures of coalition entry for other institutions have also to be devised to confirm the

generalization of our findings as well as their limits.

19GAM is already a consolidated methodology in statistics. The foundational text on GAM (Hastie andTibshirani, 1990) has had over than 9,740 cites on Google Scholar until January, 2014.

26

References

Abreu, Marcelo de Paiva. 1994. O Brasil na Rodada Uruguai do Gatt: 1982-1993. In Temas

de Polıtica Externa Brasileira II, ed. Gelson Fonseca Junior e Sergio Henrique Nabuco de

Castro. Sao Paulo: Paz e Terra pp. 325–352.

Albala-Bertrand, J. M. 1993. “Natural Disasters Situations and Growth: A Macroeconomic

Model for Sudden Disaster Impacts.” World Development 21(9):1417–1434.

Alesina, Alberto F, Stelios Michalopoulos and Elias Papaioannou. 2012. “Ethnic inequal-

ity.”.

Alvarez, Mike, Jose Antonio Cheibub, Fernando Limongi and Adam Przeworski. 1996.

“Classifying political regimes.” Studies in Comparative International Development

31(2):3–36.

Angrist, Joshua D. and Jorn-Steffen Pischke. 2008. Mostly Harmless Econometrics: An

Empiricist’s Companion. 1 ed. Princeton University Press.

Baccini, Leonardo. 2012. “Cheap Talk: Transaction Costs, Quality of Institutions, and

Trade Agreements.” European Journal of International Relations 0(0):1–38.

Baier, Scott L and Jeffrey H Bergstrand. 2007. “Do free trade agreements actually increase

members’ international trade?” Journal of international Economics 71(1):72–95.

Baron, Reuben M and David A Kenny. 1986. “The moderator–mediator variable distinction

in social psychological research: Conceptual, strategic, and statistical considerations.”

Journal of personality and social psychology 51(6):1173.

Barton, John H., Judith L. Goldstein, Timothy E. Josling and Richard H. Steinberg. 2008.

The Evolution of the Trade Regime: Politics, Law, and Economics of the GATT and the

WTO. Princeton University Press.

Beck, Nathaniel and Simon Jackman. 1998. “Beyond linearity by default: Generalized

additive models.” American Journal of Political Science 42(2):596–627.

Bhagwati, Jagdish N. 1990. “Departures from multilateralism: Regionalism and aggressive

unilateralism.” The Economic Journal 100(403):1304–1317.

27

Bhagwati, Jagdish N and Hugh T Patrick. 1990. Aggressive unilateralism: America’s 301

trade policy and the world trading system. Univ of Michigan Pr.

Bown, Chad P. 2005. “Participation in WTO dispute settlement: Complainants, interested

parties, and free riders.” The World Bank Economic Review 19(2):287–310.

Braithwaite, John. 2004. “Methods of power for development: weapons of the weak,

weapons of the strong.” Mich. J. Int’l L. 26:297.

Burges, Sean. 2013. “Mistaking Brazil for a Middle Power.” Journal of Iberian and Latin

American Research 19(2):286–302.

Busch, Marc L, Eric Reinhardt and Gregory Shaffer. 2009. “Does legal capacity matter?

A survey of WTO Members.” World Trade Review 8(04):559–577.

Cameron, A. Colin and Pravin K. Trivedi. 1998. Regression Analysis of Count Data (Econo-

metric Society Monographs). Cambridge University Press.

Cameron, A. Colin and Pravin K. Trivedi. 2005. Microeconometrics: Methods and Appli-

cations. Cambridge University Press.

Chamberlin, John. 1974. “Provision of collective goods as a function of group size.” The

American Political Science Review 68(2):707–716.

Cheibub, Jose Antonio, Jennifer Gandhi and James Raymond Vreeland. 2010. “Democracy

and dictatorship revisited.” Public Choice 143(1-2):67–101.

Chen, Xi and William D Nordhaus. 2011. “Using luminosity data as a proxy for economic

statistics.” Proceedings of the National Academy of Sciences 108(21):8589–8594.

Chipman, Hugh A, Edward I George and Robert E McCulloch. 2010. “BART: Bayesian

additive regression trees.” The Annals of Applied Statistics 4(1):266–298.

Coase, Ronald Harry. 1960. “The Problem of Social Cost.” The Journal of Law and Eco-

nomics 3:1–44.

Collier, David and Robert Adcock. 1999. “Democracy and dichotomies: A pragmatic

approach to choices about concepts.” Annual Review of Political Science 2:537–565.

28

Cooper, Andrew Fenton, Richard A Higgott and Kim Richard Nossal. 1993. Relocating

middle powers: Australia and Canada in a changing world order. University of British

Columbia Press.

Costantini, Valeria, Riccardo Crescenzi, Fabrizio De Filippis and Luca Salvatici. 2007. “Bar-

gaining coalitions in the WTO agricultural negotiations.” The World Economy 30(5):863–

891.

Crespo Cuaresma, Jesus, Jaroslava Hlouskova and Michael Obersteiner. 2008. “Natural

disasters as creative destruction? Evidence from developing countries.” Economic Inquiry

46(2):214–226.

Dahlman, Carl J. 1979. “The problem of externality.” Journal of law and economics

22(1):141–162.

Dai, Xinyuan. 2002. “Political regimes and international trade: The democratic difference

revisited.” American Political Science Review 96(1):159–166.

Downs, George W, David M Rocke and Peter N Barsoom. 1996. “Is the good news about

compliance good news about cooperation?” International Organization 50(3):379–406.

Drahos, Peter. 2003. “When the weak bargain with the strong: negotiations in the World

Trade Organization.” International Negotiation 8(1):79–109.

Edward, Mansfield, Helen Milner and Peter Rosendorff. 2002a. “Why democracies cooper-

ate more: Electoral control and international trade agreements.” International Organi-

zation 56(3):477–513.

Edward, Mansfield, Helen Milner and Peter Rosendorff. 2002b. “Why democracies cooperate

more: Electoral control and international trade agreements.” International Organization

56(3):477–513.

Elkins, Zachary. 2000. “Gradations of democracy? Empirical tests of alternative conceptu-

alizations.” American Journal of Political Science 44:193–200.

Esteban, Joan. 2001. “Collective action and the group size paradox.” American political

science review 95(3):663–672.

Evans, Peter B. 1989. “Declining hegemony and assertive industrialization: US-Brazil

conflicts in the computer industry.” International Organization 43(2):207–38.

29

Evans, Peter B, Harold Karan Jacobson and Robert D Putnam. 1993. Double-edged diplo-

macy: International bargaining and domestic politics. Vol. 25 Univ of California Press.

Finnemore, Martha and Stephen J Toope. 2001. “Alternatives to “legalization”: richer

views of law and politics.” International Organization 55(03):743–758.

Gelman, Andrew. 2007. Data analysis using regression and multilevel/hierarchical models.

Cambridge University Press.

Gilligan, Michael J. 2010. “The transactions costs approach to international institutions.”

New York University: Unpublished Manuscript.

Gilpin, Robert and Jean M Gilpin. 1987. The political economy of international relations.

Princeton University Press Princeton.

Goldstein, Judith, Miles Kahler, Robert O Keohane and Anne-Marie Slaughter. 2000. “In-

troduction: Legalization and world politics.” International organization 54(3):385–399.

Guzman, Andrew T and Beth A Simmons. 2005. “Power Plays and Capacity Constraints:

The Selection of Defendants in World Trade Organization Disputes.” The Journal of

Legal Studies 34(2):557–598.

Hamilton, Colleen and John Whalley. 1989. “Coalitions in the Uruguay Round.”

Weltwirtschaftliches Archiv 125(3):547–56.

Hardin, Russell. 1982. Collective Action (Rff Press). Baltimore, MD: Johns Hopkins Uni-

versity Press for Resources for the Future.

Hastie, T.J. and R.J. Tibshirani. 1990. Generalized Additive Models (Chapman & Hall/CRC

Monographs on Statistics & Applied Probability). 1 ed. Chapman and Hall/CRC.

Higgott, Richard A and Andrew Fenton Cooper. 1990. “Middle power leadership and

coalition building: Australia, the Cairns Group, and the Uruguay Round of trade nego-

tiations.” International Organization 44(04):589–632.

Hoekman, Bernard M. and Michel M. Kostecki. 2010. The Political Economy of the World

Trading System: From GATT to WTO. Oxford University Press.

Hsiao, Cheng. 2003. Analysis of Panel Data (Econometric Society Monographs). 2 ed.

Cambridge University Press.

30

Hurrell, Andrew. 2006. “Hegemony, liberalism and global order: what space for would-be

great powers?” International Affairs 82(1):1–19.

Imai, Kosuke, Luke Keele, Dustin Tingley and Teppei Yamamoto. 2011. “Unpacking the

black box of causality: Learning about causal mechanisms from experimental and obser-

vational studies.” American Political Science Review 105(4):765–789.

Kennedy, Peter. 2008. A Guide to Econometrics. 6 ed. Wiley-Blackwell.

Keohane, Robert O. 1969. “Lilliputians’ Dilemmas: Small States in International Politics.”

International Organization 23(02):291–310.

Keohane, Robert O. 1989. International Institutions And State Power: Essays In Interna-

tional Relations Theory. Westview Press.

Lafer, Celso. 1998. A OMC e a regulamentacao do comercio internacional: uma visao

brasileira. Livraria do Advogado Editora.

Leamer, Edward E. 1985. “Sensitivity analyses would help.” The American Economic

Review 75(3):308–313.

Leeds, Brett Ashley. 1999. “Domestic Political Institutions, Credible Commitments, and

International Cooperation.” American Journal of Political Science 43(4):979–1002.

Levi, Margaret and Gillian H Murphy. 2006. “Coalitions of contention: The case of the

WTO protests in Seattle.” Political Studies 54(4):651–670.

MacCullagh, Peter and John Ashworth Nelder. 1989. Generalized linear models. Vol. 37

CRC press.

Mansfield, Edward D and Eric Reinhardt. 2008. “International Institutions and the Volatil-

ity of International Trade.” International Organization 62(04):621–652.

Mansfield, Edward D and Helen V Milner. 2010. “Regime Type, Veto Points, and Prefer-

ential Trading Arrangements.” Stanford Journal of International Law 46:219.

Mansfield, Edward D, Helen V Milner and B Peter Rosendorff. 2000. “Free to trade:

Democracies, autocracies, and international trade.” American Political Science Review

94(2):305–321.

31

Mansfield, Edward D, Helen V Milner and B Peter Rosendorff. 2002. “Replication, realism,

and robustness: Analyzing political regimes and international trade.” American Political

Science Review 96(1):167–170.

Mansfield, Edward D, Helen V Milner and Jon C Pevehouse. 2008. “Democracy, veto

players and the depth of regional integration.” The World Economy 31(1):67–96.

Marshall, Monty G and Keith Jaggers. 2002. “Polity IV project: Political regime charac-

teristics and transitions, 1800-2002.”.

Martin, Lisa L. 2000. Democratic commitments: Legislatures and international cooperation.

Princeton University Press.

McGillivray, Fiona and Alastair Smith. 2008. Punishing the prince: a theory of interstate

relations, political institutions, and leader change. Cambridge University Press.

Meunier, Sophie. 2000. “What single voice? European institutions and EU-US trade nego-

tiations.” International organization 54(1):103–135.

Miguel, Edward, Shanker Satyanath and Ernest Sergenti. 2004. “Economic shocks and civil

conflict: An instrumental variables approach.” Journal of political Economy 112(4):725–

753.

Murray, Michael P. 2006. “Avoiding Invalid Instruments and Coping with Weak Instru-

ments.” The Journal of Economic Perspectives 20(4):111–132.

Narlikar, Amrita. 2003. International Trade and Developing Countries: Bargaining Coali-

tions in GATT and WTO. Routledge.

Narlikar, Amrita and Diana Tussie. 2004. “The G20 at the Cancun Ministerial: Developing

countries and their evolving coalitions in the WTO.” The World Economy 27(7):947–966.

Narlikar, Amrita, Martin Daunton and Robert M Stern. 2012. The Oxford Handbook on

the World Trade Organization. Oxford University Press.

North, Douglass C. 1990. Institutions, institutional change and economic performance.

Cambridge University Press.

Odell, John S. 2009. “Breaking deadlocks in international institutional negotiations: the

WTO, seattle, and doha.” International Studies Quarterly 53(2):273–299.

32

Olson, Mancur. 1965. The logic of collective action: public goods and the theory of groups.

Harvard University Press.

Olson, Mancur and Richard Zeckhauser. 1966. “An economic theory of alliances.” The

Review of Economics and Statistics 48(3):266–279.

Ostrom, Elinor. 1990. Governing the commons: The evolution of institutions for collective

action. Cambridge University Press.

Patel, Mayur. 2007. “New faces in the green room: Developing country coalitions and

decision-making in the WTO.” Global Economic Governance Programme Working Paper

33:2007.

Polachek, Solomon W. 1997. “Why democracies cooperate more and fight less: the relation-

ship between international trade and cooperation.” Review of International Economics

5(3):295–309.

Preeg, Ernest H. 1995. Traders in a brave new world: The Uruguay round and the future

of the international trading system. University of Chicago Press.

Przeworski, Adam, Michael E Alvarez, Jose Antonio Cheibub and Fernando Limongi. 2000.

Democracy and development: political institutions and well-being in the world, 1950-1990.

Cambridge University Press.

Raffer, Kunibert and Hans Wolfgang Singer. 2001. The Economic North-South Divide: Six

Decades of Unequal Development. Edward Elgar Publishing.

Ricupero, Rubens. 1998. Integration of developing countries into the multilateral trading

system. In The Uruguay Round and Beyond: Essays in Honour of Arthur Dunkel, ed.

Jagdish Bhagwati and Mathias Hirsch. Ann Arbor, MI: University of Michigan Press.

Rose, Andrew K. 2004. “Do We Really Know That the WTO Increases Trade?” The

American Economic Review 94(1):98–114.

Sattler, Thomas and Thomas Bernauer. 2011. “Gravitation or discrimination? Deter-

minants of litigation in the World Trade Organisation.” European journal of political

research 50(2):143–167.

Schirm, Stefan A. 2010. “Leaders in need of followers: Emerging powers in global gover-

nance.” European Journal of International Relations 16(2):197–221.

33

Schott, Jeffrey J. 2000. The WTO after Seattle. Peterson Institute.

Sell, Susan K and Aseem Prakash. 2004. “Using ideas strategically: The contest between

business and NGO networks in intellectual property rights.” International Studies Quar-

terly 48(1):143–175.

Simmons, Beth A. 1998. “Compliance with international agreements.” Annual Review of

Political Science 1(1):75–93.

Singh, JP. 2006. “Coalitions, developing countries, and international trade: Research find-

ings and prospects.” International Negotiation 11(3):499–514.

Skidmore, Mark and Hideki Toya. 2002. “Do natural disasters promote long-run growth?”

Economic Inquiry 40(4):664–687.

Sobel, Michael E. 1982. “Asymptotic confidence intervals for indirect effects in structural

equation models.” Sociological methodology 13(1982):290–312.

Sovey, Allison J. and Donald P. Green. 2011. “Instrumental Variables Estimation in Political

Science: A Readers’ Guide.” American Journal of Political Science 55(1):188–200.

Staiger, Douglas and James H Stock. 1997. “Instrumental variables regression with weak

instruments.” Econometrica 65(3):557–586.

Steinberg, Richard H. 2002. “In the shadow of law or power? Consensus-based bargaining

and outcomes in the GATT/WTO.” International Organization 56(2):339–374.

Subramanian, Arvind and Shang-Jin Wei. 2007. “The WTO promotes trade, strongly but

unevenly.” Journal of International Economics 72(1):151–175.

Tingley, Dustin, Teppei Yamamoto, Luke Keele and Kosuke Imai. 2014. “mediation: R

Package for Causal Mediation Analysis.” Journal of Statistical Software. Forthcoming .

Tomz, Michael, Judith L Goldstein and Douglas Rivers. 2007. “Do we really know that the

WTO increases trade? Comment.” The American Economic Review 97(5):2005–2018.

Tussie, Diana, ed. 2009. The Politics of Trade: The Role of Research in Trade Policy and

Negotiation. Republic of Letters.

Veiga, Pedro Da Motta. 2005. Brazil and the G20 group of developing countries. Managing

the Challenges of WTO Participation. Cambridge University Press.

34

Williamson, Oliver E. 1979. “Transaction-cost economics: the governance of contractual

relations.” Journal of Law and Economics 22(2):233–261.

Wolfe, Robert. 2006. “New Groups in the WTO Agricultural Trade Negotiations: Power,

Learning and Institutional Design.”.

Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data. 2

edition ed. The MIT Press.

Wooldridge, Jeffrey M. 2012. Introductory Econometrics: A Modern Approach. 5 ed.

Cengage Learning.

Yu III, Vicente Paolo B. 2008. “Unity in Diversity: Governance Adaptation in Multilateral

Trade Institutions Through South-South Coalition-Building.”.

35

Table 1: Basic Information on Coalitions—coaltrade and coalwto

Coalitions Creation Number of Countries Main Arena

African Group 1997 41 GATT/WTOAfrican, Caribbean and Pacific (ACP) Group 2001 56 Non-GATT/WTOAir Transport Services 1986 9 GATT/WTOAPEC 1989 14 Non-GATT/WTOCafe au Lait (G-20 + G-9) 1983 27 GATT/WTOCairns Group 1986 15 GATT/WTOCaricom 1997 12 Non-GATT/WTOColorado Group or Friends of Trade Facilitation 2004 40 GATT/WTOCore Group on Singapore Issues 2001 13 GATT/WTOCore Group on Trade Facilitation 2005 22 GATT/WTOCotton-4 or West and Central African countries 2003 4 GATT/WTOEuropean Union (EU) 1993 12 Regional IntegrationFIPs plus Friends 2005 35 GATT/WTOFive Interested Parties (FIPs) 2004 29 GATT/WTOFood Importers’ Group (FIG) 1986 7 GATT/WTOFriends of Ambition in NAMA 2006 8 GATT/WTOFriends of Anti-Dumping Negotiation 2003 10 GATT/WTOFriends of Environmental Goods 2005 37 GATT/WTOFriends of Fish 1998 7 GATT/WTOFriends of Geographical Indications 1998 9 GATT/WTOFriends of the Development Box 1999 14 GATT/WTOG-10 (1) 1982 9 GATT/WTOG-10 (2) 2003 10 GATT/WTOG-11 2005 11 GATT/WTOG-20 (1) 1983 18 GATT/WTOG-20 (2) 2003 22 GATT/WTOG-24 on Services 1999 23 GATT/WTOG-4 or ”New Quad” 2005 28 GATT/WTOG-4 plus Japan 2005 4 GATT/WTOG-6 2005 5 GATT/WTOG-90 2003 16 GATT/WTOG-33 or Friends of Special Products 2003 46 GATT/WTOGroup of developing countries 2005 19 GATT/WTOInformal Group of Developing Countries (IGDC) 1986 9 GATT/WTOJoint Proposal Group 2005 16 GATT/WTOLeast-developed countries (LDC) group 1999 7 GATT/WTOLike-Minded Developing Countries on Mode 4 1996 8 GATT/WTOLike-Minded Group (LMG) 1996 8 GATT/WTOMERCOSUR 1991 3 Regional IntegrationMiddle Group 2005 34 GATT/WTONAFTA 1992 3 Regional IntegrationNAMA-11 2005 10 GATT/WTONon-Trade Concerns 2000 8 GATT/WTOParadisus Group 2000 6 GATT/WTOParagraph 6 countries 2004 11 GATT/WTOQuad 1994 15 GATT/WTOQuint 1989 16 GATT/WTORecently Acceded Members (RAMs) 2003 13 GATT/WTOServices Core Group 2005 10 GATT/WTOSmall and Vulnerable Economies (SVEs) 1996 11 GATT/WTOSmall Island Developing States 1991 5 GATT/WTOSmall Vulnerable Coastal States 2003 8 GATT/WTOTRIPS/Public Health Coalition 2001 60 GATT/WTOTropical Products Group 2005 12 GATT/WTO“W52” SPONSORS 2008 108 GATT/WTO

Coalitions’ names are in the first column. Their creation dates are shown in the second column. In the thirdcolumn, we included the number of countries. In the last column, we shown their main areas (GATT/WTO,Non-GATT/WTO, and Regional Integration). The variable “coaltrade” includes all the coalitions in ourdataset. The variable “coalwto” includes only GATT/WTO coalitions. There are two different coalitionsnamed both G-10 and G-20.

36

Table 2: Descriptive Statistics for the Main Variables

Hemisphere Variable N Mean SD

Coalition Entry (coaltrade) 3189 0.318 0.745Coalition Entry (coalwto) 3189 0.286 0.714Log (real GDP) 2853 24.52 2.081

World Democracy (Dichotomous) 3181 0.601 0.490Democracy (Polity) 2819 3.437 7.232Log (Open Trade) 2998 4.196 0.617Number of Natural Disasters 3189 2.031 3.607Lag 7 (Natural Disasters) 3189 1.530 3.039

Coalition Entry (coaltrade) 2392 0.309 0.741Coalition Entry (coalwto) 2392 0.280 0.717Log (real GDP) 2126 24.651 2.171

North Democracy (Dichotomous) 2384 0.661 0.473Democracy (Polity) 2067 3.847 6.772Log (Open Trade) 2236 4.229 0.616Number of Natural Disasters 2392 2.006 3.863Lag 7 (Natural Disasters) 2392 1.525 3.272

Coalition Entry (coaltrade) 797 0.344 0.757Coalition Entry (coalwto) 797 0.305 0.704Log (real GDP) 727 24.136 1.739

South Democracy (Dichotomous) 797 0.422 0.494Democracy (Polity) 752 2.311 8.269Log (Open Trade) 762 4.103 0.609Number of Natural Disasters 797 2.107 2.697Lag 7 (Natural Disasters) 797 1.547 2.197

37

Table

3:

First

Stage,

IV-2S

LS

,an

dO

LS

Regression

s

First

Sta

geO

LS

IV-2S

LS

First

Stag

eO

LS

IV-2S

LS

First

Stage

OL

SIV

-2SL

SF

irstS

tageO

LS

IV-2

SL

S(1

)(2)

(3)(4)

(5)(6)

(7)(8)

(9)(10)

(11)(12)

A:DependentVaria

ble:co

altra

de

Log

(Real

GD

P)

0.631

1.3520.626

1.414

0.7031.531

0.5021.498

(0.072

)***

(0.401)**(0.0

71)***(0.434)**

(0.082)***(0.477)**

(0.076)***

(0.735)*L

ag7

(Natu

ral

Disa

sters)0.038

0.035

-0.168

0.0330.022

(0.008

)***

(0.007)***

(0.139)

(0.006)***(0.006)**

Dem

ocra

cy(D

ichoto

mou

s)0.214

0.016(0.051

)***(0.064)

Dem

ocra

cy(P

olity

)0.021

-0.001-0.021

(.004)***(0.004)

(0.012)+L

og(O

pen

Tra

de)

0.4280.314

-0.148(0.070)***

(0.073)***(0.362)

F-sta

tistic21.62

23.4524.89

12.65H

ausm

antest

(Pro

b)

0.0000.000

0.0000.0

00O

bserva

tion

s2,8

532,8

532,853

2,845

2,8452,845

2,4962,496

2,4962,853

2,8532,853

Nu

mb

erof

cou

ntries

144

144

144144

144

144

130130

130144

144144

B:DependentVaria

ble:co

alw

toL

og(R

ealG

DP

)0.6

191.390

0.6201.4

780.705

1.6000.482

1.554(0

.069)**

*(0.403)**

(0.070)***

(0.438)**(0.081)***

(0.470)**(0.0

71)***(0.739)*

Dem

ocracy

(Dich

otom

ou

s)-0

.014

-0.214

(0.059)

(0.136

)D

emocracy

(Polity

)-0.004

-0.025(0.004)

(0.012)*L

og(O

pen

Tra

de)

0.331-0.167

(0.070)***

(0.363)H

au

sman

test(P

rob

)0.000

0.0000.000

0.000

Ob

servation

s2,8

532,853

2,845

2,845

2,4962,496

2,8532,853

Nu

mb

erof

cou

ntries

144

144144

144130

130144

144

GD

PR

eal,

Lag

of

7yea

rsof

the

Num

ber

of

Natu

ral

Disa

sters,D

emocra

cyand

Log

of

Op

enT

rade

first

stage,

linea

rreg

ression

(OL

S),

and

instru

men

tal

varia

ble

two-sta

ge

least

square

regressio

ns

(IV-2

SL

S).

F-sta

tisticco

rresponds

toth

etest

of

the

null

hyp

oth

esisth

at

the

coeffi

cient

on

the

exclu

ded

instru

men

teq

uals

0.

The

Hausm

an

testco

mpares

two

estimato

rsw

here

one

isco

nsisten

tunder

both

the

null

and

the

altern

ativ

ehyp

oth

esis,w

hile

the

oth

eris

consisten

tonly

under

the

null.

The

null

hyp

oth

esisis

rejectedif

the

two

estimato

rsare

dissim

ilar.

All

specifi

catio

ns

inclu

de

country

fixed

effects

and

hav

eb

ootstra

psta

ndard

errors

clustered

at

the

country

level.

Bootstra

psta

ndard

errors

inparen

thesis.

+p<

.1,

*p<

.05,

**

p<

.01,

***

p<

.001.

38

Table

4:

Log

itan

dN

egat

ive

Bin

omia

lR

egre

ssio

ns

Log

itN

egat

ive

Bin

omia

lL

ogit

Neg

ativ

eB

inom

ial

Log

itN

egat

ive

Bin

omia

lL

ogit

Neg

ati

ve

Bin

om

ial

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

A:DependentVariable:co

altra

de

Log

(rea

lG

DP

)1.

836

0.09

41.

801

0.08

71.

843

0.13

81.

580.2

3(0

.229

)***

(0.0

42)*

(0.2

35)*

**(0

.043

)*(0

.257

)***

(0.0

5)**

(0.3

08)*

**(0

.046)*

**D

emocr

acy

(Dic

hot

om

ous)

0.14

30.

164

(0.2

74)

(0.1

62)

Dem

ocr

acy

(Poli

ty)

0.02

20.

018

(0.0

18)

(0.0

11)+

Log

(Op

enT

rad

e)0.

541

0.9

8(0

.343

)(0

.173)*

**

Ob

serv

atio

ns

2,77

52,

775

2,76

72,

767

2,43

32,

433

2,77

52,7

75

Nu

mb

erof

cou

ntr

ies

137

137

137

137

124

124

137

137

B:DependentVariable:co

alw

toL

og(r

eal

GD

P)

2.12

80.

072

2.09

70.

0766

2.16

80.

038

1.82

50.2

06

(0.2

32)*

**(0

.053

)(0

.241

)***

(0.0

54)

(0.2

5)**

*(0

.054

)(0

.316

)***

(0.0

6)**

Dem

ocr

acy

(Dic

hot

om

ous)

0.13

8-0

.102

(0.2

71)

(0.1

83)

Dem

ocr

acy

(Poli

ty)

0.01

70.

005

(0.0

2)(0

.012

)L

og(O

pen

Tra

de)

0.64

0.9

95

(0.3

59)+

(0.1

96)*

**

Ob

serv

atio

ns

2,77

52,

775

2,76

72,

767

2,43

32,

433

2,77

52,7

75

Nu

mb

erof

cou

ntr

ies

137

137

137

137

124

124

137

137

GD

PR

eal,

Dem

ocr

acy

and

Log

of

Op

enT

rade

logis

tic

and

neg

ati

ve

bin

om

ial

regre

ssio

ns.

All

spec

ifica

tions

incl

ude

countr

yfixed

effec

tsand

hav

eb

oots

trap

standard

erro

rscl

ust

ered

at

the

countr

yle

vel

.B

oots

trap

standard

erro

rsin

pare

nth

esis

.+

p<

.1,

*p<

.05,

**

p<

.01,

***

p<

.001.

39

Table

5:

OL

S,

Logit

and

Negative

Bin

omial

Regression

s

OL

SO

LS

Logit

Logit

Negative

Bin

omial

Negativ

eB

inom

ial

(1)(2)

(3)(4)

(5)(6)

A:DependentVaria

ble:co

altra

de

Log

(real

GD

P)

0.5060.571

1.5651.67

0.230.275

(0.074)***(0.086)***

(0.312)***(0.341)***

(0.047)***(0.05)***

Dem

ocra

cy(D

ichoto

mou

s)-0.04

0.077-0.025

(0.07)(0.281)

(0.185)D

emocra

cy(P

olity

)-0.004

0.020.002

(0.004)(0.018)

(0.012)L

og

(Op

enT

rad

e)0.321

0.3020.529

0.3450.975

0.965(0.077)***

(0.084)***0.343

(0.343)(0.17)***

(0.162)***O

bserva

tion

s2,845

2,4962,767

2,4332,767

2,433N

um

ber

of

cou

ntries

144130

137124

137124

B:DependentVaria

ble:co

alw

toL

og

(real

GD

P)

0.4910.553

1.8131.911

0.2310.192

(0.07)***(0.08)***

(0.321)***(0.336)***

(0.062)***(0.063)**

Dem

ocra

cy(D

ichoto

mou

s)-0.075

0.063-0.328

(0.065)(0.282)

(0.216)D

emocra

cy(P

olity

)-0.008

0.013-0.014

(0.004)+(0.02)

(0.0143)L

og

(Op

enT

rad

e)0.346

0.3440.636

0.5271.042

1.104(0.074)***

(0.079)***(0.36)+

(0.356)(0.196)***

(0.184)***O

bserva

tion

s2,845

2,4962,767

2,4332,767

2,433N

um

ber

of

cou

ntries

144130

137124

137124

Full

linea

r(O

LS),

logistic

and

neg

ativ

ebin

om

ial

regressio

nm

odels.

All

specifi

catio

ns

inclu

de

country

fixed

effects

and

hav

eb

ootstra

psta

ndard

errors

clustered

at

the

country

level.

Bootstra

psta

ndard

errors

inparen

thesis.

+p<

.1,

*p<

.05,

**

p<

.01,

***

p<

.001.

40

Figure

1:

GA

TT

/W

TO

Coal

itio

nM

emb

ers

from

1982

to20

08by

Nor

th-S

ou

thH

emis

ph

eric

Div

isio

n

All

GA

TT

/W

TO

mem

ber

shav

ejo

ined

coaliti

ons

at

least

once

from

1982

to2008,

soth

at

coaliti

on

entr

yand

GA

TT

/W

TO

mem

ber

ship

over

laps.

Countr

ies

shaded

inbla

ckare

South

ern

countr

ies—

bel

owth

eeq

uato

rline.

Countr

ies

shaded

ingre

yare

Nort

her

nco

untr

ies—

ab

ove

the

equato

rline.

Non-s

haded

countr

ies

wer

enot

GA

TT

/W

TO

mem

ber

sunti

l2008.

41

Figure 2: Multilevel (mediation) models with varying intercepts by year (1994—2004)

(a) Average Causal Mediation Effect(ACME), Average Direct Effect (ADE)

and Total Effect.

(b) Average Causal Mediation Effect(control) with varying intercept by

year.

(c) Average Causal Mediation Effect(treated) with varying intercept by

year.(d) Average Direct Effect (control)

with varying intercept by year.

(e) Average Direct Effect (treated)with varying intercept by year.

(f) Total Effect with varying interceptby year.

Marginal effects of the log of real GDP on coalition entry (coaltrade) mediated by the number of delegatesin Geneva from 1994 to 2004 with the following controls: democracy, log of trade openness and North-Southhemispheres were used as control variables. The model presents a dichotomous measure of democracy. Dotsrepresent the point estimates. Horizontal lines depict 95% confidence intervals (CIs). A vertical line at zeroare displayed to make the interpretation of our CIs clearer.

42

Figure 3: Multilevel (mediation) models with varying intercepts by year (1994—2004)

(a) Average Causal Mediation Effect(ACME), Average Direct Effect (ADE)

and Total Effect.

(b) Average Causal Mediation Effect(control) with varying intercept by

year.

(c) Average Causal Mediation Effect(treated) with varying intercept by

year.(d) Average Direct Effect (control)

with varying intercept by year.

(e) Average Direct Effect (treated)with varying intercept by year.

(f) Total Effect with varying interceptby year.

Marginal effects of the log of real GDP on coalition entry (coalwto) mediated by the number of delegates inGeneva from 1994 to 2004 with the following controls: democracy, log of trade openness and North-Southhemispheres were used as control variables. The model presents a dichotomous measure of democracy. Dotsrepresent the point estimates. Horizontal lines depict 95% confidence intervals (CIs). A vertical line at zeroare displayed to make the interpretation of our CIs clearer.

43

Figure 4: Multilevel (mediation) models with varying intercepts by year (1994—2004)

(a) Average Causal Mediation Effect(ACME), Average Direct Effect (ADE)

and Total Effect.

(b) Average Causal Mediation Effect(control) with varying intercept by

year.

(c) Average Causal Mediation Effect(treated) with varying intercept by

year.(d) Average Direct Effect (control)

with varying intercept by year.

(e) Average Direct Effect (treated)with varying intercept by year.

(f) Total Effect with varying interceptby year.

Marginal effects of the log of real GDP on coalition entry (coaltrade) mediated by the number of delegatesin Geneva from 1994 to 2004 with the following controls: democracy, log of trade openness and North-Southhemispheres were used as control variables. The model presents a continuous measure (Polity) of democracy.Dots represent the point estimates. Horizontal lines depict 95% confidence intervals (CIs). A vertical lineat zero are displayed to make the interpretation of our CIs clearer.

44

Figure 5: Multilevel (mediation) models with varying intercepts by year (1994—2004)

(a) Average Causal Mediation Effect(ACME), Average Direct Effect (ADE)

and Total Effect.

(b) Average Causal Mediation Effect(control) with varying intercept by

year.

(c) Average Causal Mediation Effect(treated) with varying intercept by

year.(d) Average Direct Effect (control)

with varying intercept by year.

(e) Average Direct Effect (treated)with varying intercept by year.

(f) Total Effect with varying interceptby year.

Marginal effects of the log of real GDP on coalition entry (coalwto) mediated by the number of delegates inGeneva from 1994 to 2004 with the following controls: democracy, log of trade openness and North-Southhemispheres were used as control variables. The model presents a continuous measure (Polity) of democracy.Dots represent the point estimates. Horizontal lines depict 95% confidence intervals (CIs). A vertical lineat zero are displayed to make the interpretation of our CIs clearer.

45

Figure 6: Generalized Additive Models (GAMs)—(coaltrade)

9 10 11 12 13

0.2

0.4

0.6

0.8

log(Real GDP)

Est

imat

ed v

alue

s (2

.46

df)

(a) Marginal effects of the log of real GDP on coalitionentry (coaltrade)

1.0 1.5 2.0 2.5

0.2

0.4

0.6

0.8

log(Openness to Trade)

Est

imat

ed v

alue

s (3

.89

df)

(b) Marginal effects of the log of trade openness oncoalition entry (coaltrade)

−10 −5 0 5 10

0.2

0.4

0.6

0.8

Polity Score

Est

imat

ed v

alue

s (4

.15

df)

(c) Marginal effects of democracy (Polity IV) on coalitionentry (coaltrade)

Effects of the log of real GDP (6a), of the log of trade openness (6b) and of the Polity IV (6c) on coalitionentry (coaltrade). Vertical axes (y) display the values of coalition entry. Y-axes scales were transformedby a Logit link function to represent probability intervals. Each graph shows the estimated value of thedegrees of freedom for each smoothing function. A 95% confidence interval for the cubic smoothing splinesis shaded.

46

Figure 7: Generalized Additive Models (GAMs)—(coalwto)

9 10 11 12 13

0.2

0.4

0.6

0.8

log(Real GDP)

Est

imat

ed v

alue

s (3

.21

df)

(a) Marginal effects of the log of real GDP on coalitionentry (coalwto)

1.0 1.5 2.0 2.5

0.2

0.4

0.6

0.8

log(Openness to Trade)

Est

imat

ed v

alue

s (3

.42

df)

(b) Marginal effects of the log of trade openness oncoalition entry (coalwto)

−10 −5 0 5 10

0.2

0.4

0.6

0.8

Polity Score

Est

imat

ed v

alue

s (4

.62

df)

(c) Marginal effects of democracy (Polity IV) on coalitionentry (coalwto)

Effects of the log of real GDP (7a), of the log of trade openness (7b) and of the Polity IV (7c) on coalitionentry (coaltrade). Vertical axes (y) display the values of coalition entry. Y-axes scales ere transformed by aLogit link function to represent probability intervals. Each graph shows the estimated value of the degrees offreedom for each smoothing function. A 95% confidence interval for the cubic smoothing splines is shaded.

47