1

FREE TRADE AGREEMENTS AND THE CONSOLIDATION OF

DEMOCRACY*

Xuepeng Liu

Kennesaw State University

Emanuel Ornelas

London School of Economics

September 2010

Abstract

We develop a model to study the relationship between participation in free trade agreements (FTAs) and

the sustainability of democracy. We find that, because of the rent-destructing effect of FTAs, they can help

democracy to “consolidate” in a country. If authoritarian groups seek power largely to appropriate rents, an FTA

reduces their incentives to do so, increasing the likelihood that democracy will endure in the country. In turn, this

implies that governments in fledgling, unsecure democracies will have an extra motive to engage the country in

FTAs: to strengthen democracy and, if a democratic reversal seems inevitable, to constrain the rent-seeking

activities of future autocrats. In a dataset with 133 countries over 1948-2007, we find strong empirical support for

our theoretical predictions.

INCOMPLETE and PRELIMINARY

COMMENTS & SUGGESTIONS VERY WELCOME

* We thank Kishore Gawande, Alan Spearot, and Benjamin Zissimos for useful comments on an earlier version of this

paper.

2

1. INTRODUCTION

Over 200 free trade agreements have been formed since the early 1990s, according to the

World Trade organization. This is also the period where the “third wave of democratization,” as

defined in the political science literature (see Huntington 1991), has taken place. We argue that these

two phenomena, depicted in Figure 1, are not independent from each other. Specifically, we claim

that participation in free trade agreements (FTAs) serves as a commitment device to destroy future

protectionist rents. Since such rents are attractive for autocratic groups, FTAs lower their incentives

to seek power. While this has little value in established democracies, where the rule of law is strong

and the risk of authoritarian disruption is negligible, it can be of great importance for new, unstable

democracies. Trade gains (or losses) aside, these threatened states should therefore be particularly

keen to seek involvement in FTAs. This can help to explain the seemingly puzzling outbreak of

regionalism since the early 1990s, while also helping to rationalize the third democratic wave.

We are not the first to claim that participation in trade agreements is correlated to being

democratic, but to our knowledge we are the first to rationalize, and qualify, this relationship.

Policymakers, for example, often cite political factors such as the “strengthening of democracy” as a

central force behind integration initiatives.1 While it is generally difficult to distinguish true

intentions from rhetoric, political scientists evaluating this issue empirically have indeed found a

positive correlation between participation in FTAs and democracy.2 What is missing is a theory that

allows us to understand this relationship better and test it appropriately. It could be, for example,

that other factors induce countries to form FTAs and to become democratic, but causality does not

follow. In contrast, it could be that FTA participation induces countries to become democratic, or

vice-versa. Clearly, a theoretical model is necessary to guide an empirical evaluation of this

relationship.

We provide this theoretical basis by extending the regional integration model developed by

Ornelas (2005) to allow for endogenous changes in the political regime. In that otherwise standard

model, at any trade regime domestic firms exchange protection for contributions with the

1 For instance, in the 2001 summit congregating the potential signatories of the Free Trade Area of the

Americas, “President Bush said striking down trade barriers was critical to sustaining democracy […] throughout the region” (New York Times, 4/18/2001). To the extent that can be inferred from public speeches, the other region’s leaders shared a very similar view. Similarly, the demand of Eastern and Central European countries for membership in the European Union has often been linked to the countries’ democratic concerns.

2 For example, Mansfield, Milner and Rosendorff (2003) find that pairs of democratic countries are more likely to create trade agreements than pairs in which at least one country has an authoritarian political regime. Mansfield, Milner and Pevehouse (2008) show further that this holds for different types of trade agreements except the “shallowest” ones, according to their five-tier classification. Conversely, Pevehouse (2002) finds that participation in international organizations (which include mainly free trade agreements but also other international organizations) tends to increase the longevity of new democracies.

3

government, which cares about national welfare and the contributions it receives. A similar game

takes place at the ex ante stage, when the government defines the trade regime—whether or not to

form an FTA. Key to understand the impact of an FTA is the recognition that the equilibrium of the

(ex post) external tariff game changes with the constraint imposed by the agreement on the internal

tariffs. Taking this into account, the main message that arises from the analysis is that, even though

an FTA still permits lobbying for protection against excluded countries, the volume of protectionist

rents falls with the formation of the agreement. As a result, only arrangements that improve welfare

enough to compensate for the lower rents are politically viable.

In a dynamic setting this implies that, all else equal, groups motivated mainly by office rents

will have lower incentives to seek power if the country is engaged in an FTA (and withdrawal from

the agreement is costly). Motivated by empirical findings from the received literature, we consider

that authoritarian groups tend to fit this description best.3 This is also plausible theoretically. After

all, due to the aptitude of authoritarian administrations to resort to violence rather than to rely on

accountability to keep power, their incentives to pursue policies that favor the population at large

tend to be lower than they would be under a democratic administration.4 But if the gain of

authoritarian groups from keeping power are lower if the country is engaged in an FTA, while the

costs and risks from attempting a coup d’état are unaltered by the agreement, the likelihood of

democratic failure will, all else being equal, be lower if the country participates in an FTA.

Now, if the incumbent government in an unstable democracy realizes this effect of

“democratic consolidation,” it will tend to seek participation in FTAs more actively than it would

otherwise. The reason is two-fold. First, the FTA will weaken the authoritarian threat. Second, even

if the dictatorial group takes control despite the FTA, the agreement will constrain its rent-extraction

activities. Hence, unstable democracies will tend to enter in FTAs more frequently than other

countries, all else being equal. In turn, participation in FTAs will tend to increase the likelihood of

democracy survival in those countries.

3 Nalin and Torstensson (1995) find that dictatorships are more likely than democracies to pursue

distortionary redistributive policies. Specifically to trade policies, Banerji and Ghanem (1997), Mansfield, Milner and Rosendorff (2000) and Rama (1994) present evidence that authoritarian regimes are associated with increased trade protection and trade regulations. Mitra, Thomakos and Ulubasogly (2002), estimating welfare and contributions’ weights for Turkey in periods of both democratic and authoritarian administrations, provide additional support for the presumption that, relative to special interests, welfare concerns are more important for democracies than for dictatorships. In a recent paper, Aidt and Gassebner (2008) find evidence of more protectionist policies in autocratic states with respect to both formal (tariffs) and informal (‘red tape’) trade policies.

4 See for example Dixit (2008) for an equivalent assumption, which he uses when analyzing differences in the policymaking process under autocracy and democracy when rulers have to rely on bureaucrats to implement policies.

4

We look for, and find, empirical support for our theoretical results by analyzing the

formation of FTAs and the strength of democracy in 133 countries over 1948-2007. On the one hand,

we find that greater participation in FTAs lowers the likelihood of democratic failure in a country.

On the other hand, we find that a high risk of democratic breakdown induces countries to

participate in more FTAs. Our findings are robust to different measures of FTAs and the

consideration of endogeneity.

Interestingly, the predictions hold well for full-fledged FTA and customs unions, signed

according to the GATT Article XXIV. When partial-scope preferential trade agreements (PTAs) are

included, the results remain significant in most cases but are usually somewhat weaker. This is what

we would expect, as those PTAs, signed under the Enabling Clause of the GATT, allow for many

exceptions in regional trade liberalization.

One of our empirical challenges is to define how unstable a democracy is. We do so by

relying on a recent study by Persson and Tabellini (2009), who estimate the likelihood of democratic

breakdown employing the concept of “democratic capital.” Democratic capital has two components,

one domestic and one foreign. The domestic component takes into account the history of democracy

in the country. The longer the country has experienced democracy, and the more recent is its

democratic experience, the greater the country’s current stock of domestic capital. The concept of

foreign capital encompasses instead current levels of democracy in the rest of the world. The greater

the number of democratic countries, and the closer those democratic nations are to a certain country,

the greater is that country’s current stock of foreign capital. Along with other covariates, these two

components of democratic capital allow us to estimate the likelihood of democratic failure in a

country. In the sense that we find that participating in a greater number of FTAs significantly

reduces this probability, regionalism helps to “consolidate democracies.”

Having estimated the likelihood of democratic failure, we use its fitted values to estimate

entry in/formation of FTAs. In doing so, we consider only the portion of the likelihood that is not

predicted by FTA participation. Our finding that higher levels of regime uncertainty does indeed

induce democratic governments to seek participation in FTAs helps to rationalize the regionalism

wave since the early 1990s, when many countries became democratic but, due to their limited

experience with democracy, faced significantly high levels of political instability.

The paper proceeds as follows. Section 2 describes the model. Section 3 presents the analysis

of the incentives to form a free trade agreement. Section 4 discusses our empirical strategy. The data

is presented in Section 5. We show our results in Section 6. We conclude in Section 7.

5

2. THE MODEL

2.A. The economic structure

We consider a 3-country, N-sector competitive economy where in each sector there is a

“natural importer” country that would import the good from the other two countries under free

trade. Goods are produced under constant returns to scale. One unit of the numeraire good 0 is

produced with one unit of labor. All other goods j = 1…N – 1 are produced with labor and a sector-

specific factor. Thus, whenever good 0 is produced in equilibrium, which we assume to be the case,

the wage rate equals unity and all general equilibrium forces are absorbed by that sector.

The analysis is conducted from the perspective of a “Home” country. Home’s population

consists of a continuum of agents with measure one. Each agent is endowed with one unit of labor,

whereas specific factors are owned by a negligible fraction of the population. Consumers have

quasi-linear utility of the form [ ]∑−

=−+= 1

1

2 2/)(N

j

jj0 qAqqU , which generates demand D j = A – p j

for good j.

Home is the natural importer of goods m = 1…M, country Y is the natural importer of a

subset E of different goods, and country Z is the natural importer of the remaining (N – M – E – 1)

non-numeraire products. Home’s owners of the specific factor used in sector j earn πj(pj), where pj

denotes the price of good j in Home’s market. The domestic supply of each imported good m is

Sm(pm) = dmpm and the supply of each exported good x is Sx(px) = dxpx, where dx > dm > 0. An

analogous specification applies for the supply and demand structures of countries Y and Z. Home

can use specific import tariffs in each import sector; other policy instruments are assumed

unavailable. We represent Home’s tariff on imports from country j by tj, j = Y, Z. Because all import

sectors are identical, we will write prices and tariffs without sector-identifying superscripts.

Prices in the three countries are linked by arbitrage conditions. For a generic product

imported by Home, this condition is

(1) p = pY + tY = pZ + tZ,

provided that tariffs are not prohibitive. Using this arbitrage condition, market-clearing requires

(2) D(p) – Sm(p) = Sx(p – tY) – D(p – tY) + Sx(p – tZ) – D(p – tZ).

Using the expressions for demand and supplies defined above, condition (2) can be rewritten as

(3) ρ++γ= )(),(ˆ YZYZ ttttp ,

where γ ≡ 3A/(3+dm+2dx) and ρ ≡ (1+dx)/(3+dm+2dx).

When Home is member of a free trade agreement, it follows GATT’s requirement of non-

6

discrimination. When Home is in an FTA, imports from the FTA partner are duty free, but imports

from the excluded country remain taxed, although the country’s external tariff will change as a

result of the FTA, as we will see.

2.B. The political structure

There is a democratically elected government that rules Home. The group represented in the

government enjoys power because there are rents for holding office. The sources of those rents are

transfers/bribes offered by the private sector, which benefit government officials in exchange for

more favorable policies. Thus, the rents are specific to incumbency, as for example in Besley and

Coate (2001).

The group in power cares also about national welfare. As in the literature of strategic debt

issuance, we consider that a party’s welfare concerns reflect the links with its “constituency.” In

general, the larger the constituency of the elected government, the higher the weight the government

attaches to national welfare, relative to office rents. A direct implication is that a larger constituency

induces the government to internalize a greater share of the distortions created by its policies. This

specification therefore presumes that the welfare concerns of a political group, stemming from the

group’s link with its constituency, are unrelated with incumbency, unlike the transfers obtained

through interactions with the private sector.

Let us define the measures of welfare. Welfare generated in an import sector is denoted by

Wm(t), whereas Wx represents welfare from an export sector. The former is defined as the sum of

consumers’ surplus, tariff revenue and producers’ surplus generate in that sector; the latter is

defined as the sum of consumers’ and producers’ surplus in the sector.5 Welfare aggregated across

all non-numeraire import and export sectors is then WM(t) ≡ MWm(t) and WX ≡ (N - M - 1)Wx,

respectively. National welfare, W(t), aggregates welfare across all sectors:

∑∑−

+==++=++≡ 1

11)(1)(1)(

N

Mx

xM

m

mXM WtWWtWtW .

The preference of the political party in office—the government—is specified as

(4) ,),(),(1

11 ∑∑−

+==+≡ N

Mx

xmM

m

m GTtGTtG

with Gx ≡ Wx/b and

5 Note that we denote welfare in import-competing sectors as a function of the tariff, but not in export

sectors. In reality, Wx also depends on tariffs, but on those imposed by foreign countries Y and Z. Since those tariffs are given from the perspective of the Home government under any trade regime, we employ this briefer representation for notational ease.

7

(5) ,)(1

),( mmmm TtWb

TtG +≡

where Tm denotes the transfer from import-competing sector m to the government, ,1∑ =

≡ M

m

mTT and

b is a parameter that is inversely related to the size of the constituency of the government. We refer

to b as the “political bias” of the government. Thus, if the government’s constituency were very

small, the government’s political bias would be very high and it would care mainly about rents.

Conversely, if the government’s constituency were very large, the government’s political bias would

be very low and it would largely disregard rents.

We assume producers within each industry can overcome free-riding problems and act

jointly in their lobbying activities. Because of the symmetry and independence across sectors, we

focus on a single import-competing sector. The net payoff of producers in such a sector corresponds

to the industry’s aggregate profits, πm(t), subtracted of the transfers it gives to the local government,

Tm.

As in Maggi and Rodríguez-Clare (1998), we model the interaction between government and

each domestic industry as a Nash bargaining game, where each side obtains half of the total surplus

from the negotiations process. Under the Nash bargaining protocol, the outcome of the bargaining

process is jointly efficient. Thus, the “political tariff” resulting from this interaction satisfies

(6) )]()(max[arg tbtWt mmp π+= ,

where the term in brackets represents (up to a constant) the joint payoff of the government and the

industry in a representative import-competing sector. To simplify exposition, we restrict the analysis

to the case where the solution of problem (6) is interior. This corresponds to assuming that b < bmax ≡

(1+dm)(dx–dm)/(1+dx)dm.

2.C. Equilibrium payoffs

In the absence of lobbying activities, the government can do no better than set the tariff in

each import sector to maximize national welfare. As a result, its payoff in that case would be given

by Wm(tp(b=0))/b, or simply Wm(b=0)/b. Analogously, the domestic industry m would obtain a

payoff of πm(b=0) ≡ πm(tp(b=0)).

We can then define the “political rents” created in the lobbying process in each import-

competing sector as

(7) ( ) ( )[ ])0()0()()(1 =π+=−π+≡ bbbWbbbWb

PR mmmmm ,

8

where functions Wm and πm are evaluated at the political tariff when they are represented as a

function of b. The expression in the first parenthesis of (7) is the maximized joint payoff of the

government and the industry (multiplied by b), while the expression in the second parenthesis is the

value of the same function in the absence of lobbying. The difference between these two expressions

represents the surplus that the lobbying process adds to the joint payoff of the government and the

industry.

In equilibrium, the government obtains its reservation payoff in the industry plus its share of

the political rents:

(8) 2

)0(1 m

mm PRbW

bG +== .

Aggregating across all sectors, we can then write (4) evaluated at the equilibrium as

(9) 2

1)0(

1),(

PRW

bbW

bTtGG XMp ++==≡ ,

where .1∑ =

≡ M

m

mPRPR Hence, the government obtains in equilibrium its reservation utility,

[WM(b=0) + WX]/b, plus half of the political rents. This makes clear that the group in power does

not internalize the welfare distortions due to its use of the political tariff.

In contrast, if the same political group were out of power, its payoff would be different even

if the tariff were the same. The reason is that the group does not receive any rents if it is not in a

position to enact policies. Accordingly, in that case the group would receive none of the available

political rents, and its equilibrium payoff H would reflect only the concerns for its constituency:

(10) Xinc

Mp Wb

bWb

tGH11

0 +=≡ )(),( ,

where binc denotes the political bias of the political group in office. Since WM(b=0) ≥ WM(b i n c) and PR ≥

0, it follows directly from (9) and (10) that there are benefits from holding office.

2.D. Coup threat

We consider a simple 2-period environment where there is a group of citizens representing a

segment of the population that is not represented in power but may attempt to take power through

force, initiating a dictatorship in the country. We are not specific about which segment of the

population is represented by this group; it could be the military as well as part of the country’s

capitalists or the upper class, for example. In any case, whenever a coup is attempted and is

successful, the authoritarian group takes power in the second period.

9

We assume the potential dictatorial group cares predominantly about its own well-being.

That is, rather than considering the potential dictatorship as “benevolent,” as sometimes assumed in

economic analyses, we treat it as “kleptocratic.” Moreover, since the supporting group of a

dictatorship tends to be considerably smaller than that of a democratically elected government, the

forces limiting rent-seeking behavior tend to be weaker in a dictatorship. This suggests that an

autocratic government would attach a higher weight on rents vis-à-vis national welfare than the

democratic government. Accordingly, we consider that

A1: bD ≤ bA,

where we use identifiers D (democratic) and A (autocratic) to distinguish between variables related

to the incumbent democratic government and the authoritarian group, respectively.

We let the probability of success of a coup depend negatively on the country’s stock of

“democratic capital.” The notion of democratic capital (DC) is introduced and developed by Persson

and Tabellini (2009), and corresponds to a measure of the strength of the country’s democratic

institutions. In nations with enduring democratic tradition, where the rule of law is strong,

democratic capital tends to be abundant, virtually precluding the possibility of political disruption.

Conversely, in countries lacking solid institutions, where the rule of law is weak, democratic capital

is likely to be scarce, thus opening a tangible opportunity for successful coups. In the definition of

Persson and Tabellini (2009), an important element shaping the stock of DC in a country is the

country’s democratic history. In fact, as North (1990) points out, the costs of altering political

institutions are very low when they are new, but increase as they get older.

We assume that the precise level of democratic capital is not known with certainty, but its

expected value, ,DC is. Specifically, we define the democratic capital of the Home country as

,= DCDC θ where θ is a random variable with expected value E(θ) = 1 distributed accordingly to

distribution function Φ in the interval [0, 2].

The probability of success of a coup depends also on how the population reacts to the

attempt. In particular, success will be more likely the stronger the “support” of the pro-coup citizens

and the weaker the “resistance” from the segments of the population opposed to the coup. We call

the difference between support and resistance in the population the “net support” for a coup, and

denote it by s. We normalize the units of s and DC so that an attempted coup is successful if and

only if s > DC. Thus, from the perspective of the group considering subverting the country’s

democratic order, the probability of success is

10

(11) p ≡ prob (s ≥ DC) = prob (s/DC ≥ θ) = Φ(s/DC ),

where p denotes the probability of success of the coup.

Both the democratic government and the authoritarian group discount future payoffs

according to a (common) discount factor δ, where δ ∈ [0, 1]. To understand when the authoritarian

group will attempt to subvert the country’s democratic order, we model the group’s problem as

simply as possible. In particular, we assume that, if the takeover attempt is successful, the

authoritarian group imposes a dictatorship in the country and obtains its office payoff GA in the

second period. If the takeover attempt is unsuccessful, the group bears a fixed cost K > 0.6

When a coup is attempted, the present value payoff of the incumbent and of the

authoritarian group are represented, respectively, as

(12) ])[( DDDD pHGpG +−δ+=Γ 1

and

(13) ]))([( AAAA pGKHpH +−−δ+=Γ 1 .

In no coup were attempted, the incumbent government and the authoritarian group receive,

respectively, GD and HA in each period.

The (risk-neutral) authoritarian group attempts to take power if and only if the expected

utility from the endeavor is positive: ΓA > (1 + δ)HA. Using (13), this condition is equivalent to

(14) KpHGp AA )()( −>− 1 .

That is, the authoritarian group will attempt to take power if its expected gain from seeking power is

large relative to the expected cost of a failed coup.

To make explicit what is behind the probability of success of the coup and the gains from

holding office, we use the definition of p and expressions (9) and (10) to rewrite condition (14) as

(15) [ ] KDCs

DCsPRbWbW

b

A

DMM

A )/(

)/(-1

2)(-)0(

1

ΦΦ

>+= .

In a consolidated democracy, where either the support for a coup is very low or democratic capital is

knowingly very high, an attempt against the county’s democratic system is unlikely, unless the costs

of failure are too low—which is rarely the case—or the gains from holding power are very

significant. Our central goal is to analyze how a free trade agreement affects the latter, and through

that channel the endurance of democracy in a country.

6 Parameter K provides a proxy for the many kinds of penalties that could apply in such a case—incarceration, extradition, death and the like.

11

Naturally, an FTA can be used to affect future policies only if its reversal is costly enough to

inhibit withdrawal from the arrangement by future governments. While here we simply assume that

FTAs are irreversible, Ornelas (2007) presents a straightforward extension of a similar model where,

paralleling McLaren (2002), governments incur in “negotiating costs” when forming (or

withdrawing from) FTAs. That extension, which treats irreversibility as an equilibrium result, shows

that the rationale developed in this paper requires only the existence of a cost to reverse established

FTAs.7

3. THE DECISION TO FORM A FREE TRADE AGREEMENT

A free trade agreement between two countries is represented by the elimination of tariffs on

each other’s imports. Thus, the equilibrium under an FTA is entirely analogous to the one described

in Section 2, the only difference being the constraint imposed on the partner’s reciprocal import

tariffs. Without loss of generality, we let Home’s potential FTA partner be country Y.

An FTA can be implemented by the incumbent government for reasons related or unrelated

to the authoritarian threat. There are four possibilities. First, it is possible that the country is already

a consolidated democracy, in the sense that condition (15) is not satisfied regardless of the existence

of FTAs. This is the standard case analyzed in the regionalism literature, and it is not our goal to

analyze it further in this paper. Rather, we focus on situations where FTAs are formed for “strategic”

reasons.

Second, it is possible that the country’s democracy is so fragile that condition (15) is satisfied

whether or not there is an FTA in place. In that case, while an FTA cannot be used to prevent a coup,

the possibility of a reversal to a dictatorship can affect the incentives of the incumbent government

with respect to the formation of the agreement.

Finally, it is possible that an FTA affects the expected payoff of the authoritarian group and,

as a result, its incentives to attempt to take power. In general, an FTA could either increase or

decrease the incentives for a coup, making it worth seeking when it would not be without the

agreement or vice-versa, making the coup not worth pursuing when it would be in the absence of

7 It is worth noting that irreversibility is coherent with history, as preferential trading arrangements de facto

implemented are seldom turned down later on. Even in the rare circumstances when authoritarian regimes gained control of a country that participated in an effective trade agreement, the arrangement has been honored, as for example in Swaziland, a member of SACU. The only exception to this rule seems to be the Andean Pact, from which President Hugo Chávez withdraw Venezuela in 2005. The other cases of implemented agreements being later disrupted took place in Central America (CACM) and in the Caribbean (CARIFTA/CARICOM), but in both cases they were disrupted due to balance of payments constraints during the debt crisis of the 1980’s. Both were fully reactivated in the early 1990’s.

12

the agreement.

We begin analyzing the cases where the FTA is not pivotal for the decision of the

authoritarian group—i.e. when condition (15) is unaffected by the FTA. To do so, we need first to

describe the effects of an FTA on the level of available political rents and the role of the political

parameter b in shaping the welfare effects of an FTA. These results set the basis for the analysis of

the political viability of FTAs.

3.A. The rent destruction effect

Ornelas (2005) shows that an FTA moderates the role of political economy forces in the

determination of tariffs, and that the mitigation of the politically driven distortions corresponds to a

source of welfare gain that is more relevant, the more far-reaching the government’s political

motivations. Furthermore, an FTA diminishes the rents created in the lobbying process. Intuitively,

because the arrangement provides free access to the partner’s exporters, the market share of the

domestic industry shrinks, at any given external tariff. As a result, the FTA makes any price increase

brought by a marginal increase in the external tariff less valuable for the import-competing

industries, lowering their incentives to lobby for higher external tariffs. In equilibrium, these lower

incentives imply less lobbying, a lower external tariff and fewer rents for the government. The

following lemma summarizes these effects.

Lemma 1. The rent destruction effect of FTAs (Ornelas 2005)

Everything else constant, an FTA

(a) improves Home’s welfare by more (or reduces it by less), the higher the government’s

political bias; and

(b) reduces the political rents generated in the political process, this reduction being larger,

the higher the government’s political bias.

Lemma 1 allows us to analyze the conditions under which the Home government would

choose to form an FTA.8 The decision regarding the formation of an FTA is based on the anticipated

impact of the agreement. The government implements the agreement if and only if it increases the

government’s present value payoff.

8 Naturally, an FTA is formed only if all prospective members endorse it. We conduct the discussion from

the perspective of the Home country, but an analogous analysis would apply for country Y.

13

Before proceeding to analyze how the possibility of political disruption affects the

willingness of the democratic government to form free trade agreements, it is important to note that,

under A1, the authoritarian group will not want to form an FTA if the democratic government has

rejected it. On the other hand, if the authoritarian group is willing to form a certain FTA, the

democratic government will want to do too, strategic motivations notwithstanding. To show this, we

henceforth attach subscript “F” to all variables when they are evaluated under an FTA. We also

adopt subscript “∆F” to represent the equilibrium change in any variable due to the FTA. For

example, XFW∆ denotes the aggregate welfare change in the export sectors due to the agreement,

whereas )( DMF bW∆ and )0=(bW M

F∆ denote, respectively, the aggregate welfare impact of the FTA on

the import sectors under the ruling of the incumbent democratic government and under an

administration whose only concern is national welfare (or equivalently, if lobbying were effectively

banned).

Lemma 2. Strategic considerations aside, if the democratic government does not want to form an

FTA, the authoritarian would not want to form it either. However, if the authoritarian group

would want to implement an FTA, the democratic government wants to implement it, too.

Proof: In the absence of strategic considerations (i.e. if there were no possibility of political

disruption), the democratic government wants to implement an FTA if .0>DFG∆ From (9), this

condition is equivalent to

(16) DFD

XF

MF PRbWbW ∆∆∆ >+= -])0([2 .

Similarly, if in power, the authoritarian group would be willing to form the FTA if

(17) AFA

XF

MF PRbWbW ∆∆∆ >+= -])0([2 .

Now notice that, under A1, part (b) of Lemma 1 implies that the right-hand side of (17) is greater

than the right-hand side of (16). Therefore, if condition (17) is satisfied, condition (16) is also

satisfied. Conversely, if condition (16) is not satisfied, condition (17) is not satisfied either. �

Since the authoritarian group would not support an agreement that the democratic

government ordinarily rejects, it follows that, if the democratic government decided to back an FTA

only for strategic reasons, the FTA would not be supported by the authoritarian group in case it

gains power (since the authoritarian group has no further strategic motivation once it gains power).

3.B. FTAs that do not affect the probability of political disruption

14

We begin analyzing the case where there is a possibility of political disruption but this

possibility is unaffected by the existence of FTAs.

The equilibrium payoff of the incumbent democratic government under the FTA

corresponds to

(18) ]+)-1[(+= DF

DF

DF

DF pHGpG δΓ .

in this case. Thus, the condition under which the democratic government supports the FTA when

the authoritarian threat is inevitable is 0>ΓΓ≡Γ∆DD

FDF - . Using equations (12) and (18), D

F∆Γ can be

rewritten as

]+)1[(+= DF

DF

DF

DF pHGpG ∆∆∆∆ δΓ - .

Using expressions (9) and (10) and manipulating, this expression becomes

(19)

δ++δ+

+=δ+=Γ ∆∆∆

∆∆XFA

MF

DF

DMF

D

DF WbpW

PRbbWp

b)1()(

2)0()]1(1[

1- .

Thus, the incumbent democratic government supports the FTA in this case if

(20) 0)1(2)(2])0(2)][-1(1[ >δ++δ++=δ+ ∆∆∆∆XFA

MF

DFD

MF WbpWPRbbWp .

We know that ,0>∆XFW since the preferential treatment under the FTA improves Home’s terms of

trade vis-à-vis the two other countries in the E sectors where Home exports to country Y. On the

other hand, )0( =∆ bW MF < 0 and 0<D

FPR∆ by Lemma 1.9

The interesting case is when the democratic government changes its stance toward an FTA

because of the authoritarian threat. An FTA is (ordinarily) politically feasible if

(21) 0])0([2 >++= ∆∆∆DFD

XF

MF PRbWbW .

The next proposition shows that the authoritarian threat can make an FTA politically feasible even

when condition (21) does not hold.

Proposition 1. Even if the authoritarian threat cannot be affected, the mere possibility of political

disruption can turn an otherwise politically unfeasible FTA into a viable one. By contrast, the

possibility of disruption cannot render unfeasible an otherwise feasible FTA.

Proof: We need to show first that DF∆Γ increases with p. Using (19), we have that

(22)

=δ=Γ ∆

∆∆∆

2)0()(

DF

DMFA

MF

D

DF PR

bbWbWbdp

d-- .

9 When b = 0, the government chooses tariffs to maximize welfare in the import sectors. Since the FTA

constrains the tariffs on imports from Y to zero, it must reduce welfare in those sectors when b = 0.

15

We know from Lemma 1 that 0<DFPR∆ and that the welfare impact of an FTA is increasing in the

political bias of the government, so that .0>)0=()( bWbW MFA

MF ∆∆ - Accordingly, expression (22) is

unambiguously positive, so DF∆Γ increases as the probability of disruption rises. As a result, an FTA

that is politically unfeasible when there is no chance of political disruption can become viable if the

likelihood of political disruption increases enough. That is, an FTA that does not satisfy condition

(21) can satisfy criterion (20) for sufficiently high p. On the other hand, the reverse cannot happen: if

an FTA is politically viable when there is no chance of political disruption, it remains feasible if a

possibility of change in power through force arises. That is, an FTA that satisfies condition (21)

satisfies criterion (20) for any p > 0. �

Proposition 1 shows that the possibility of political disruption can enhance the political

feasibility of FTAs by creating a “strategic” motivation for their adoption. Strategically supported

FTAs arise when, among conditions (20) and (21), only the former is satisfied, so that

(23) )0(0)0( >Γ<≤=Γ ∆∆ pp DF

DF .

An FTA can be implemented for strategic reasons because the democratic government, if out of

power, will not receive any of the lobbying-related rents. In that case, the government would benefit

from an FTA because the agreement constrains the welfare-distorting political activities of the

authoritarian group if it gets in power. Thus, a government that expects to lose power to a dictatorial

group might seek an FTA simply to constrain the policies of the incoming authoritarian group. Since

this strategic motivation is more relevant when disruption is more likely, it follows that “political

instability” tends to incite the formation of free trade agreements.

The political biases of both the incumbent government and of the authoritarian group affect

the possibility of strategically supported FTAs. In particular, higher political biases enlarge the scope

for this type of arrangements. The same is true for the number of Home’s import-competing sectors,

M.

Proposition 2. The set of parameters under which the possibility of political disruption can turn an

FTA politically viable increases with the political biases of the democratic (bD) and the

autocratic (bA) groups, as well as with the size of Home’s import-competing sector (M).

16

Proof: To prove these results, it suffices to show that the probability of disruption, p, is a strategic

complement of bA, bD and M in the function ,DFDb ∆Γ which gives the criterion for the political

viability of FTAs. This function is represented in the left-hand side of (20). We have that

0)()(2 >δ=Γ ∆∆

A

AMF

A

DFD

db

bdW

dpdb

bd, and

0)()(2 >+δ=Γ ∆

∆∆

D

DF

DDF

D

DFD

db

dPRbPR

dpdb

bd

2 - ,

where all inequalities follow from Lemma 1. From the definition of welfare aggregated across all

non-numeraire import sectors, and of aggregated political rents, we also have that ( )DmF

DF PRMPR ,= ∆∆

and )]0=()([=)0=()( bWbWMbWbW mFA

mF

MFA

MF ∆∆∆∆ -- . Therefore,

0)0()()( ,

2

>

=δ=Γ∆∆∆

∆ DmF

DmFA

mF

DFD PR

bbWbW

dpdM

bd

2-- .

Hence, the set of parameters under which condition (23) is satisfied enlarges as bA, bD and M

increase. �

The intuition behind Proposition 2 is as follows. The more biased toward special interests is

the authoritarian group, the larger the distortions it would create if it held power. In that case, the

role of FTAs in moderating distortions and enhancing welfare is magnified (Lemma 1), so a larger bA

makes the incumbent democratic government more inclined to use an FTA to limit the rent-seeking

activities of the potential dictatorship. Now, if the democratic government is itself very receptive to

the politically generated rents, it will in general be sign a rent-destructing FTA. This anti-FTA force

is partially neutralized, however, if the incumbent believes its chance of keeping power is small,

since in that case the loss of rents brought about by the FTA would be born mainly by the

authoritarian group. If the weakening of this effect is sufficiently strong, the incumbent may find it

worthwhile to support the agreement.

Finally, if Home were small relative to its FTA partner, in the sense of having a relatively

large non-numeraire import-competing sector, the incentives of the democratic government to form

the agreement because of the authoritarian threat are greater as well. If Home imports more

intensely from its partner, the agreement is more rent destructing. While this is helpful for the

country as a whole, it is detrimental to those in office who benefit from those rents. Under the threat

of political disruption, however, the government understands that the loss of rents will be borne

17

instead by the authoritarian group, if it is successful in gaining power. The destruction of rents is

therefore less critical in the democratic government’s evaluation of the agreement.

As indicated in the Introduction, the idea that governments can manipulate state variables to

constrain their successors’ choices was first advanced in the macroeconomics political economy

literature. Prominent examples in that line of research are the pioneering contributions of Alesina

and Tabellini (1990) and Persson and Svensson (1989), who employ such a rationale to study the

politics of debt issuance. A similar reasoning is employed here to show that a democratic

government, when faced with the prospect of political disruption, may want to limit the ability of

the potential authoritarian government to create rents through interactions with the domestic

industry. We show that an FTA can be an effective tool for that purpose.

3.C. FTAs that can help secure democracies

The analysis above considers the case where a free trade agreement is not pivotal in the

decision of the authoritarian group to attempt to take power through force. But this need not be the

case. That is, an FTA may change the sign of condition (15). We show that an FTA can indeed change

the sign of condition (15), but the change can go in only one direction. Specifically, an FTA can

prevent a coup from happening, but it cannot provoke a coup that would not occur in the absence of

the agreement.

Proposition 3. If the authoritarian group does not intend to initiate a coup in the absence of trade

agreements, an FTA cannot induce it to initiate one. On the other hand, the formation of a

sufficiently rent destructing FTA can free the country from the authoritarian threat. This is

more likely to happen, the greater the political bias of the democratic government (bD) and

the size of Home’s non-numeraire import-competing sector (M).

Proof: In the absence of trade agreements, the authoritarian group attempts to take power with a

coup if condition (15) is satisfied. With an FTA, a similar condition applies:

(24) [ ] KDCs

DCsPRbWbW

b

AF

DMF

MF

A )/(

)/(-1

2)(-)0(

1

ΦΦ

>+= .

Clearly, the only difference between conditions (15) and (24) is in the expressions’ left-hand sides,

which denote the gains of the authoritarian group from getting power. On the other hand, the FTA

impacts neither the probability of success of a coup nor the costs of a failed coup attempt.

Subtracting the left-hand side of inequality (15) from the left-hand side of inequality (24), we obtain

18

(25) [ ] 02

)(-)0(1 <+= ∆

∆∆

AF

DMF

MF

A

PRbWbW

b,

where the negative sign follows directly from Lemma 1. Thus, if condition (15) is not satisfied,

condition (23) will not be satisfied either, implying that an FTA cannot provoke a coup that

otherwise would not occur. Conversely, condition (15) can be satisfied while condition (23) is not,

implying that an FTA can be critical to prevent the authoritarian group from seeking power. The

range of parameters under which this happens is larger, the greater the number of Home’s non-

numeraire import-competing sectors, since the left-hand side of (25) decreases with M:

[ ] 02

)(-)0(1)]25([ ,

<+== ∆∆∆

AmF

DmF

mF

A

PRbWbW

bdM

lhsd.

The set of parameters under which an FTA can free the country from the authoritarian threat

increases also with the political bias of the incumbent democratic government, since the left-hand

side of (25) decreases with bD as well:

0<)]([1

-=)]25([

D

DMF

AD db

bWd

bdb

lhsd ∆,

where 0>/)]([ DDMF dbbWd ∆ by Lemma 1. �

Proposition 3 shows that, because of the rent destructing effects of FTAs, a free trade

agreement can critically reduce the incentives of the authoritarian group to attempt to subvert the

country’s democratic system. In this sense, an FTA can help to constrain the emergence of

authoritarian regimes, especially if the bloc is significantly rent-destructing, as in that case it will be

more effective in lowering the gains from power of the authoritarian group. Building on the

common notion that the availability of rents can entice political turbulence—while the unavailability

of rents can prevent it10—the proposition’s novelty stems from the recognition of free trade

agreements as instruments to restrain the gains from rent-seeking behavior.

We still need to ask, however, whether the incumbent democratic government would

actually want to implement the arrangement. The next proposition shows that the possibility of

using an FTA to block a coup necessarily raises the government’s political benefits with the

agreement.

Proposition 4. An FTA can become politically feasible by being pivotal to prevent a coup.

10 This is, for example, Olson’s (1993) main message.

19

Proof: When an FTA cannot prevent the authoritarian group from seeking power through force, it is

politically viable if .0])1[( >+δ+=Γ ∆∆∆∆DF

DF

DF

DF pHGpG - When the agreement reverses the decision of

the authoritarian group, it is adopted by the democratic government if

]+)1[(+>+ DDDDF

DF pHGpGGG -δδ ,

where the expression in the left-hand side is the present value of the government under the

agreement (and no authoritarian threat) and the right-hand side is its expected present value

without the FTA (and with the authoritarian threat). This condition can be rewritten as

(26) 0>)(+)+1( DDDF HGpG -δδ ∆ .

Now notice that the left-hand side of (26) is greater than DF∆Γ if ,DF

DF HG > which is true from the

definitions of DFG and D

FH , which are analogous to those in (9) and (10). Hence, even if DF∆Γ < 0,

condition (26) can be satisfied. �

Proposition 1 asserts that the possibility of political disruption can render feasible an

otherwise unfeasible free trade agreement. Proposition 4 indicates that the political support for an

FTA can be further enhanced if the agreement can also play a role in preventing disruption of the

political system. This is true even though here we abstract from any ideological motivation the

incumbent democratic government may have; if the government perceived a benefit per se from

maintaining democracy in the country, its incentive to form an FTA that can serve that purpose

would be further enhanced.

The result suggests that free trade agreements—especially those that are particularly

effective in destroying office rents, as those formed with larger trade partners—can be useful to

prevent the breakdown of democracy when the country does not have sufficient democratic capital

to prevent the authoritarian threat. This is often the case in nascent democracies, given the unstable

political periods that typically follow the end of dictatorial regimes. In fact, the establishment of new

democracies has often been followed by the formation of preferential arrangements (or the accession

to existing ones). This was the case, for example, of all Mercosur members, of Greece, Portugal and

Spain in their accession to the European Community, and of the EU agreements with the countries

of Central and Eastern Europe shortly after the fall of the iron curtain. The European Community

was itself established few years after the end of autocrat regimes in some of its original members

(Germany and Italy). Not surprisingly, then, the consolidation of democratic regimes is often

20

presented as one of the primary goals in the formation of free trade agreements.11 The present

analysis provides a (so far absent) coherent explanation for the link between democratic

consolidation and the establishment of free trade agreements.

4. EMPIRICAL STRATEGY

The model has two main predictions about the relationship between FTAs and democracy,

which imply the following hypotheses:

H1. Participation in FTAs lowers the likelihood of democratic failure.

H2. Unstable democracies are more likely to form FTAs.

To test H1, the dependent variable is the probability that democracy will fail in the country,

which we denote by Prob(enddemo), or the length of democratic spells. The key independent

variable is a measure of the country’s participation in FTAs.

To test H2, the dependent variable is the change in the FTA variable. The key independent

variable is a measure of unsecured democracies, which reflects the expectation that the democratic

regime may fail in the country.

As indicated in the Introduction, our problem is related to the one studied by Persson and

Tabellini (2009), who examine the determinants (in particular the effect of income) of the stability of

democracies and the impact of this perceived stability on income growth. Our empirical strategy

resembles Persson and Tabellini’s approach.

Specifically, we estimate the likelihood of democratic failure relying on the concepts of

domestic democratic capital (DOM) and foreign democratic capital (FOR) developed by Persson and

Tabellini, but adding a variable that captures participation in FTAs. DOM is a measure of the

democratic history of the country, whereas FOR measures current levels of democracy across the

world.12 Other explanatory variables include economic factors (e.g. GDP per capita, denoted by the

vector X) and geographical and institutional factors (e.g. war indicators, continent of location and

legal origin, denoted jointly by the vector Z). With the dependent variable being a dummy

11 As The Economist (4/19/2001) points out with regard to the possible creation of the Free Trade Area of the Americas,

“the elected leaders of Latin America look to the United States as an export market but also as a source of support for democracy in the region.” Similarly, when the United States first announced the intention to pursue a free trade agreement with Central America countries, there were three explicit goals with the agreement, one of which was “to support democracy in the region”—the other two were to promote U.S. exports and to advance the FTAA (www.whitehouse.gov, January 16, 2002).

12 In the next section we provide a precise definition of both variables.

21

indicating whether a democracy was interrupted (enddemo), a discrete time duration analysis such

as logit or complementary loglog (cloglog) can be implemented as follows:13

(27) Prob(enddemo) = α0 + α1FTA-1 + α2DOM-1 + α3FOR-1 + α4X-1 + α5Z-1 + u,

where subscript -1 denotes that the variable is lagged one year.14

The variable FTA in equation (27) represents a measure of the country’s degree of

participation in FTAs in a given year, or the country’s “FTA intensity.”. Measuring the FTA intensity

of a country is a non-trivial task. Most empirical regionalism papers use dummies to represent FTA

participation. While this can be adequate in other studies, such a measure is not appropriate for our

purposes. The reason is that there is wide heterogeneity in FTAs and in the countries participating in

FTAs. First, while some arrangements are fully implemented, others are not, implying few and small

preferences are actually offered. Second, a given FTA can have a very different impact on each of its

members.15 Given this, in all of our main regressions we use the share of imports from FTA

members to represent a country’s FTA intensity. This variable should vary monotonically with the

degree of implementation of the agreement and with the importance of the agreement for the

country in question. It is also a useful proxy for the degree of rent destruction of the FTA, since this

is positively associated with the share of preferential imports. To check for robustness, we also

briefly discuss when using alternative definitions for FTA intensity.

There is a concern about duration dependence in (27). If there is duration dependence, the

hazard of enddemo will depend on the duration of the democratic regime. In principle, its effect can

be either positive or negative. To capture duration dependence, we use a polynomial of a time

counter that counts the number of years passed since the beginning of the current democratic spell.

The order of the polynomial is determined by the best fit in the regressions.

Including this duration polynomial, which we denote by DUR, we rewrite our estimating

equation as

13 When the probability of positive outcomes is small, as in the case of enddemo, where democratic failures

account for only 2% in our dataset (see Table 1), a cloglog link function is similar to a logistic link function. Hence the coefficients obtained from a cloglog regression can be exponentiated and understood in terms of odds ratio. Since logistic regressions are more conventional than cloglog ones, we only report logit results. Results from cloglog regressions are very similar.

14 Some variables, such as geographic variables, legal origin and colony dummies, are constant over time. Therefore it does not matter whether they are lagged or not.

15 Consider for example NAFTA. While it has always had a very large impact on Mexico, the smallest member, its impact is much less pronounced on the United States, the largest member.

22

(28) Prob(enddemo) = α0 + α1FTA-1 + α2DOM-1 + α3FOR-1 + α4X-1 + α5Z + α6DUR + u.

A remaining concern in (28) is unobserved heterogeneity. That is, it is possible that some

countries are more likely to have interrupted democracies due to unobserved variables that are

correlated with some right-hand side variables in (18). To deal with this possibility, we also run a

random effects logit specification.16, 17

Once we have the predicted Prob(enddemo), we can use it to test our second hypothesis, that

the likelihood of democratic failure helps to explain the formation of FTAs. We also include the

economic, geographical and institutional variables used in the duration regression, except DOM and

FOR. Nothing in our theory suggests that DOM and FOR should have an independent effect on

∆FTA in addition to their indirect effects on ∆FTA through Prob(enddemo).18 This is similar to the

identification assumption of Persson and Tabellini (2009), that both components of democratic

capital affect income growth through their effects on the sustainability of democracy only. Some

may argue that the rationale to exclude FOR may not apply here, since FOR could have an

independent effect in the formation of FTAs. For this reason, we will run the ∆FTA regression with

and without FOR to check the robustness of our results.

Note also that the duration dependence terms are included in the prediction of

Prob(enddemo) but will not be included in the second stage ∆FTA regressions, which instead use

year dummies to capture the time effect. As another robustness check, we also try to include all the

regressors from the first stage enddemo regressions in the ∆FTA regression (i..e, without exclusion

variables). In this case, we rely only on the non-linearity of our first stage logit estimation to identify

the predicted hazard in our second stage ∆FTA regression.

To test H2, we then run the following specification, where we include both year dummies

and country fixed effects:19

16 No consistent fixed effect cloglog procedure is available. The fixed effect logit procedure (i.e. conditional

logit) is available but inappropriate because most of the country*spell units of observation do not experience democracy breakdown during our sample period. These observations would be dropped in fixed effect logit regressions, eliminating much of the cross-sectional variation.

17 With “frailty,” as unobserved heterogeneity is denoted in the duration literature, the logistic hazard

function becomes )]exp(1/[1 uXh tijtijt −θ−β′−+= . If we assume a normal distribution for the unobserved

heterogeneity u, we can use a random effects logit specification. 18 In fact, including them in the ∆FTA regressions has only a very small impact on the results once the

predicted Prob(enddemo) is included, with the DOM coefficient being statistically insignificant. 19 Other variables affecting FTA formation could also be included. Interestingly, even though much has been

written about regionalism, we still know little about what makes a government willing to form FTAs. Consider for example the important contribution to this topic by Baier and Bergstrand (2004). A major difference between their paper and ours is the data structure. Their data is bilateral and cross sectional so that each observation is associated

23

(29) ∆FTA = β0 + β1Prob(enddemo) -1 + β2FTA-1 + β3X-1 + β4Z-1 + v.

One may be tempted to estimate equations (27) and (29) together as a simultaneous system:

logit for equation (28) and OLS for equation (29). But this method is not applicable. First, the FTA

variable appears in levels in equation (28) but in changes in equation (29). Second, Prob(enddemo) is

unknown in equation (29).20 Therefore, we need to estimate the two equations separately.

Specifically, we replace Prob(enddemo) in equation (29) with its predicted hazard rate ( h ) from

equation (28). Because h is a function of FTAs, we do not use the variable FTA directly in the

estimation of h . Instead, we estimate h from the following regression:

(30) Prob(enddemo) = γ0 + γ2DOM-1 + γ3FOR-1 + γ4X-1 + γ5Z-1 + γ6DUR + e.

Having estimated h , we then use it in equation (29), paralleling the empirical strategy of

Persson and Tabellini (2009). We also include the squared term of h to capture possible

nonlinearities.

4.A. Duration analysis

The duration regression (logit or cloglog) we describe above is a discrete time analysis,

where we treat the time interval as discrete or grouped by year. Alternatively, we also use

continuous time duration analysis for equation (28), defining the dependent variable as the duration

of democracy spells, i.e. the number of years passed since the onset of each democracy spell.

Continuous time models include the Weibull and the Cox proportional hazard model. Hazard rates

or hazard ratios can be predicted from these regressions and then used in regression (29). We focus

on discrete-time duration analysis, but provide results from a continuous time analysis as a

robustness check.

with a country pair, while our data is a panel with country and time dimensions. Furthermore, most of the explanatory variables considered by Baier and Bergstrand are either geographical, which do not change over time, or “structural,” in the sense of changing little over time (e.g. factor endowments). Since we work with a panel, our fixed effects capture all of those fixed/almost fixed factors.

20 Simply substituting the binary variable enddemo as Prob(enddemo) in equation (29) and running logit for equation (28) would cause logical inconsistency. As summarized in Maddala (1983, p. 118), for logical consistency

either α1 or β1 in equations (28) and (29) must equal zero, which would make our study irrelevant.

24

In the Weibull model, the hazard function is 1−ϕβ′ϕ= tXth t )exp()( , where φ is the shape

parameter to be estimated and t is the duration time. The Weibull model is suitable for modeling

data with monotone hazard rates that either increase (φ >1) or decrease (φ <1) exponentially with

time (if φ=1, the Weibull distribution reduces to the exponential distribution, which is associated

with a constant hazard over time.) In the Cox model, the hazard is assumed to be

)exp()0()( tXhth β′= , where h(0) is the baseline hazard. The coefficients are assumed to be the same

regardless of group (country*spell), but the baseline hazard is allowed to be group-specific.

For the structural FTA regression, instead of using country-year panel data, we can

alternatively employ a bilateral FTA panel dataset and also apply a discrete-time duration analysis.

Each observation in the bilateral data corresponds to two countries. The dependent variable is a

dummy indicating FTA relationship between two countries in a given year. We can include in the

data the average predicted hazard rates of the two countries in the pair. We would expect to see a

positive impact of the average hazard on bilateral FTA formation.

Researchers usually resort to discrete choice regressions (e.g. probit or logit) for binary FTA

data analysis. The ordinary binary choice analysis is acceptable in cross section analysis, as in Baier

and Bergstrand (2004), if we assume most of the variations are time-invariant. A panel data analysis

is usually better as it can provide richer information over time and is more capable to handle the

problem of unobserved heterogeneity. The ordinary binary choice analysis is, however,

inappropriate for panel data analysis because it assumes that the dependent variable (FTA dummy)

is conditionally independent over time.21 Following Liu (2008, 2010), we address this problem by

applying a discrete time duration analysis. For the discrete FTA variable, this method corresponds to

standard complementary log-log (cloglog) or logistic regression. All we need to do is drop all but the

first positive outcome of the dependent variable for each country pair over the sample period. Once

the repeated “1”s are dropped, the problem of conditional dependence in the standard logit analysis

disappears. Note that the event failure in this instance is two countries agreeing to a trade agreement

and a spell is defined as the length of time until two countries form an agreement.

It is easy to show that, when the probability of positive outcomes is small as in the case of

FTAs, a cloglog link function is very close to a logistic link function. Because logit is more commonly

used than cloglog, we use logit in this paper. We need to take into account the duration dependence,

i.e. the extent to which the conditional hazard of an FTA is rising or falling over time. If year

dummies are used to account for duration dependence, all the years without any new FTA would be

21 For example, Mexico signed NAFTA with the U.S. and Canada in 1994, and NAFTA remains in place for

all the following years. So the independence assumption is obviously violated.

25

dropped because the year dummies could perfectly predict these zero outcomes. To avoid this

problem, we create a time counter to count the number of years already passed since the beginning

of our sample, as well as its square, cubic terms and so on. Then this polynomial is used to account

for the time dependence. The order of the polynomial is determined by the best fit of regressions.

We will also consider the issue of unobserved heterogeneity in the duration analysis. In the duration

analysis literature, the unobserved heterogeneity is called “frailty”, analogous to the fixed effects or

random effects in panel data. If we assume a normal distribution for the unobserved heterogeneity,

we can use random effects cloglog or random effects logit.22

4.B. Endogeneity

An important issue in our analysis is endogeneity, or reversed causality between FTA and

Prob(enddemo). Since we do not estimate the two equations jointly, we have to address the

endogeneity problem in each equation. Fortunately, this problem is avoided in both equations due

to the design of our regressions.

In our structural estimation (29), the endogeneity of Prob(enddemo) is not a problem

because we use the predicted Prob(enddemo), which is a function of exogenous variables, not

including the FTA variable. In other words, the predicted hazard in the ∆FTA equation is the

exogenous component of Prob(enddemo). In the bilateral structure regressions, the endogeneity

problem is further reduced due to the fact that we drop the repeated “1”s of the FTA dummy.

Keeping only the first positive outcome, we can hardly think of any feedback of FTAs on the

predicted hazard of enddemo. This is an additional benefit of duration analysis compared to regular

logit or probit analysis.

Regarding the endogeneity of the FTA variable in the enddemo regressions, it may not be a

problem if Prob(enddemo) only affects the changes in FTAs, not its levels. Basically, a country could

be already engaged in a couple of agreements but, for other reasons, face a high probability of

democratic failure; forming a new FTA would be a way to lower that probability. Similarly, a

country that plans to sign more FTAs (partially for the sake of rent destruction) may not necessarily

have a lot of existing FTAs. More importantly, if the risk of enddemo indeed affects FTA (positive

impact as in our structural regression), this reverse causation should most likely lead to an over-

estimation of our coefficient of the lagged FTA in the enddemo regression. Without addressing the

endogeneity issues, we find a negative effect of FTA on enddemo. Had we solved the endogeneity

22 No consistent fixed effect cloglog procedure is available. The fixed effect logit procedure (i.e. conditional

logit) is available but inappropriate due to information loss. Because most of the country pairs have never signed any FTAs, they would be dropped in fixed effect logit regressions.

26

problem, the true estimate should then be even more negative. In other words, the reverse causality

problem biases against the hypothesis that FTA reduces the risk of enddemo.

But we also provide econometric evidence that no endogeneity problem exists in our

enddemo duration regressions. To address the endogeneity problem of FTAs, we cannot use 2SLS

because our logit regression is nonlinear. We cannot use the predicted FTA from a first step

regression in the second step enddemo regression either. There is no well defined bivariate

distribution in this approach because the two steps follow normal and extreme value distributions,

respectively.

Thus, we use the two-step procedure defined as the “control function approach” developed

by Petrin and Train (2010). This procedure works as follows. In the first stage, we estimate FTA with

instrumental variables and predict the residual. In the second stage, we add to the enddemo

regression the predicted residual and possibly its higher order terms as the "control function" to

solve the endogeneity problem. The correct standard errors can be obtained by bootstrapping

methods.

For the first stage IV estimation, we use “remoteness” as our instrument. Remoteness is

defined as the distance of a country to the rest of the world weighted by all other countries’ GDP in a

given year:

(31) ∑∑≠≠

=im

mtim

mtmiit GDPGDPceDismoteness /)tan(Re .

Gravity model theories predict that remoteness, as an index of “multilateral resistance” (Anderson

and Wincoop 2003), increases bilateral trade. By similar logic, remote countries may tend to seek

FTAs to compensate for their natural economic isolation from the rest of the world. By contrast,

there is no clear theoretical reason for why remoteness could have an independent effect on the

likelihood of democratic survival (especially after controlling for FOR). Thus, we assume that

remoteness is uncorrelated with the error term in the enddemo regression. Since our remoteness

variable is time varying due to the GDP weight, it can be used as instrument in our FTA regressions

with country fixed effects. As we will see, the remoteness variable is highly significant, hence a

strong instrument for FTAs.

The control function approach treats the endogeneity as an omitted variable problem. The

predicted residual can be considered a proxy of the omitted variable. If the predicted residual as

well as its higher order terms (i.e. the control function) are jointly significant in the enddemo

27

regression, it implies that there is indeed an endogeneity problem with the variable FTA. The control

function approach can also be used to control for the omitted variables. As we will see, the control

function is highly insignificant in our estimations, implying that endogeneity is not an issue in our

enddemo regressions.

5. DATA

We have a panel with 133 countries over 1948-2007. Countries that have never experienced

democracy are dropped because we only look at democratic spells. Appendix 1 lists the countries

covered in our enddemo duration analysis. More than 200 countries are covered to constructing our

FTA measures, as explained below.

In our empirical analysis, free trade agreements cover both free trade areas and customs

unions (FTA/CU), which we call full-fledged regional trade agreements (RTAs). However, we will

check the robustness of our results by including the partial-scope RTAs as well (PTAs). Our

preferred measure of RTAs is a country’s imports with RTA partners as a share of its total imports in

a given year:

� FTA_impsh: a country’s imports from FTA/CU partners as a share of its total imports;

� RTA_impsh: a country’s imports from all RTA partners as a share of its total imports.

The import data are from the IMF Direction of Trade Statistics. We also construct some alternative

measures which will be discussed in the results section. To construct these measures, we carefully

consider the dates of the formation of new blocs, of the accession of new members, and of the de-

activation of existing blocs. Appendix 2 lists the RTAs in our dataset with their types (FTA, CU or

PTA), as well as the data sources (see footnotes).

The enddemo variable is created from Polity scores, which vary from -10 to 10, with higher

values representing more democratic regimes. We define a regime as “democratic” if its polity score

is strictly positive. For a democratic spell, enddemo is zero as long as a democracy remains

uninterrupted and becomes unity when it ends. If a democracy does not end during our sample,

enddemo only takes zeros (right-censored). These countries simply contribute a string of zeros, with

no final one, to the likelihood. There are 91 episodes of enddemo in our sample; Appendix 3 lists

them.

Notice that some countries have multiple democratic spells. In the duration analysis, we

treat those spells as if they belonged to different countries. That is, our unit observation is a

country’s democratic spell. In our duration analysis, we cluster standard errors by country*spell and

28

use country*spell random effects when considering unobserved heterogeneity across countries. The

results are similar regardless of how we deal with the issue of multiple spells. In the ∆FTA

regressions, we use country fixed effects rather than country*spell fixed effects.

The DOM and FOR variables are also created based on polity scores. DOM is defined as

∑−=

=−−=

0

0,)1(

tt

tiit dDOMτ

τ

ττδδ ,

where δ is a discount factor and di,t-τ is a dummy for a strictly positive Polity score. As Persson and

Tabellini (2009), we find that what really matters for democratic stability in DOM is current DOM

(i.e. the current democratic spell), while past DOM is usually insignificant in the regressions.

Therefore, we use current DOM (dc95_pt_cur) in all of our regressions, meaning that t0 corresponds

to the first year in which dit = 1 in the current democratic spell. For the discount factor, we adopt δ =

.95; results change little for δ ∈ [.94, .99], the range considered by Person and Tabellini.

In turn, FOR is defined as

∑≠

−=

ijt

ijjtit N

DEq

DistPolityFOR /1 ,

where Polityjt is country j’s Polity score at t (rescaled to the [0,1] interval), Distij is the distance

between the capitals of countries i and j, DEq is half the length of the equator, and Nt is the number

of independent countries in the world at t.

GDP per capita data are from the World Development Indicators (WDI) database. Data on

wars are from the Correlates of War (COW) dataset, including any extra, inter and intra wars a

country was involved. Legal origin data are from La Porta (1999). Colony history variables are from

CIA’s World Fact Book (2004). The WTO membership data are from the WTO website. Trade

openness measures are obtained the Penn World Table 6.3. Please refer to Table 1 for the definitions

of all the variables used in the regressions, as well as for some descriptive statistics. The data used

for the bilateral analysis are based on Liu (2010).

6. EMPIRICAL RESULTS

6.A. Does participation in FTAs affect democracy survival?

We study first the impact of lagged FTA participation on the duration of democracy. We

start by comparing the Kaplan-Meier nonparametric survival curves for countries with and without

FTAs. As Figure 2 shows, democracies without FTA/CU (bottom curve) fail more quickly than those

29

with FTA/CU (top curve). A logrank test also rejects the null hypothesis that both groups face the

same hazard of failure, with a p-value of 0.0014. This, however, is unconditional evidence without

control variables. In our regression analysis, we use various types of discrete-time and continuous-

time duration models.

Tables 2A and 2B show the results from duration analysis for full-fledged RTAs (FTA/CU)

and all types of RTAs, respectively. Most of the time-varying covariates are lagged by one year to

alleviate potential endogeneity problems. The coefficient of RTA import share is always negative

and significant, at least at the 10% level. This result supports our first hypothesis that FTAs can

reduce the probability of democracy failure. Regression (1) of Table 2A uses only the FTA/CU

import share variable and it is highly significant. This variable alone explains more than 5% of the

variations in enddemo as shown by the Pseudo R2. In regression (2) we add other explanatory

variables and the duration dependence terms. It turns out that a second order polynomial of the

time counter produces the best fit of the model. As expected, the duration dependence terms are

highly significant. The coefficient of duration and duration^2 are positive and negative, respectively.

This implies an inverted-U shape of the impact of duration on hazard rate with a cutoff duration of

about 23 years. In other words, when a democracy is younger than 23 years, the hazard of

democratic breakdown increases with the the age of democracy; after 23 years, the hazard starts to

decrease with the age. FTA/CU import share remains significant at the 10% level.

Due to possible unobserved heterogeneity, in regression (3) we use country*demo spell

random effects (or “frailty”). The FTA/CU import share variable again displays a negative sign, and

becomes even more significant statistically. The LR test of the random effects, however, is only

marginally significant at 10% level. Moreover, the changes in the estimated coefficients are overall

very small. This suggests that unobserved heterogeneity is relatively unimportant to explain the

breakdown of democracy.

Columns (4) and (5) of Table 2A show results for the continuous duration models, the

Weibull and the PH Cox proportional hazard models, respectively. In those models, the dependent

variable measures the age of a democracy (or the number of years passed since the onset of a

democratic regime until it was interrupted or (right-) censored).23 This variable is the same as the

time counter we use for the duration dependence in columns (2) and (3). All other covariates are the

same as those in the discrete duration analysis. The duration dependence is specified parametrically

(Weibull model) or non-parametrically (PH Cox model). The dataset is a multiple-record survival

23 If a country started with a democratic regime at the beginning of our sample (year 1948), 1948 is assumed

as the onset of the democratic spell (“left-censoring”).

30

dataset because a country may have multiple observations that are either failures or right-censored.

An ID or group variable (country*demo spell) is specified to make sure that the multiple records for

each country*demo spell are not taken as separate records for different countries. Our first

hypothesis is strongly supported by the results from the Weibull regression. The estimated shape

parameter φ is significantly larger than 1, which implies that the hazard increases with the age of

democracies. For example, this estimate implies that, after 20 years, a democracy are 24 times more

likely to fail than after 10 years, ceteris paribus (i.e., (20/10)5.6-1=24). In the PH Cox regression, the

results again support our hypothesis that more intense participation in FTAs helps to consolidate

democracies.24

In the last column of Table 2A, we rerun the enddemo regression in a linear probability

model (LPM). Despite some well-known drawbacks, LPM is simple and easy to interpret and

therefore is often used in applied work. Because LPM is inherently heterosckadastic, we include

country*demo spell fixed effects and report robust standard errors. The RTA import share variable

remains statistically significant at 5% level.

Are these estimates economically significant? The results from different models should be

interpreted differently. In the LPM, the coefficient measures the impact of the covariates on P which

denotes Prob(enddemo). In the logit model, an exponentiated coefficient provides an estimate for

the impact of a one unit change in a covariate on the odds ratio of enddemo (i.e., exp(β) = [P1/(1-

P1)]/[P0/(1-P0)], where P1 = Prob(enddemo|X1); P0 = Prob(enddemo|X0), and X0 and X1 represent the

different values of X). In the Weibull and PH Cox models, an exponentiated coefficient provides an

estimate for the impact of a one unit change in a covariate on the hazard ratio of enddemo (i.e.,

P1/P0). If the risk of democracy failure is very small, the odds P/(1-P) and hazard rate P are

approximately similar. Therefore the coefficients under logit, Weibull, and HP Cox model are more

directly comparable than those under the LPM. After some simple transformations, however, we can

approximately interpret the coefficients from any of these models in terms of changes in hazard rate

(P). The random effect logit regression result in column (3) of Table 2A implies that one percentage

point increase in FTA/CU import share will decrease the odds of enddemo (P/(1-P), similar to

hazard rate, P, when P is very small) by about 2.66% (i.e., exp(-2.697/100)-1=-2.66%). The

24 Proportional hazard (PH) refers to the effect of any covariate having a proportional and constant effect

that is invariant to when in the process the values of the covariate changes. We test the PH assumption and find no evidence that the model we specified violates the PH assumption either globally or with respect to each covariate.

31

corresponding estimates from the Weibull, PH Cox and LPM regressions are 2.35%, 2.13%, and 1.7%

decreases in the hazard.25

Other covariates usually have the expected signs. GDP/capita, current domestic capital,26

foreign capital and UK colony dummy have negative and statistically significant impacts on

enddemo, while the estimated coefficients of most other variables are insignificant in the

regressions.

Table 2B reports the enddemo results for all types of RTAs, including both the full-fledged

FTA/CU and the partial-scope PTAs. The results are similar to those in Table 2A. The coefficients of

the FTA/CU and All-RTA import share variables remain significant, except for the LPM regression.

The magnitude of the coefficients for the import share variables is however smaller than in Table 2A.

This is expected because the partial-scope RTAs are incomplete processes of preferential trade

liberalization, with many exceptions and special treatments.

6.B. Are unstable democracies more likely to seek participation in FTAs?

We now turn to our second hypothesis. We rely on a structural estimation by predicting the

hazard rate from the duration analysis and then estimating how the predicted hazard rate affects

FTA formation.

The results from our main specification are reported in the first three columns of Table 3A.

The first column shows the enddemo duration regression and the next two columns display the

∆RTA regressions with the predicted hazard for FTA/CU and All-RTA respectively. As discussed in

Section 4, we do not include the RTA variable in this first stage to avoid an endogeneity problem in

the second stage. We include a second order polynomial of the predicted hazard rate to the second

stage ∆RTA regression to check for possible nonlinear effects. In columns (2) and (3), the dependent

variable is the change in the RTA import share from the previous year. Year dummies are included

to capture the recent increasing trend in RTA growth. To control for unobserved heterogeneity

across countries, country fixed effects are also included in the regression. Finally, we exclude DOM

and FOR as well the duration dependence terms from that regression as the exclusion restriction.

Because hazard is predicted from the first stage with sampling errors, we use bootstrapping (100

replications) to adjust the standard errors in the ∆RTA regressions.

25 These estimates are in percent, not percentage. Based on the average predicted hazard of 0.023 as shown

in Table 1, these estimates amount to about 0.0006, 0.0054, 0.0005 and 0.0004 percentage point decreases on average in the hazard for xtlogit, Weibull, PH Cox and LPM respectively.

26 DOM actually bears an unexpected positive sign in the LPM, different from all the other regressions. This may indicate some potential problems with the LPM specification.

32

In column (2) FTA/CU regressions, the coefficients of the first order predicted hazard rate

and its quadratic term are both significant at 10%, positive and negative respectively. This implies an

inverted-U relationship between hazard rate and FTA formation with an estimated cutoff value of

0.225. When hazard<0.225, FTA/CU formation increases with the hazard; when hazard>0.225,

FTA/CU formation decreases with the hazard. As shown in Table 1 and Figure 3, the estimated

hazard rate ranges from 7.81E-8 to 0.379, with a mean 0.023. Only 6 of the predicted hazard rates

(among a total of 2603) are larger than the cutoff value.27 Thus, unless the risk of democratic failure

is exceedingly high, unstable democracies tend to form more FTAs than stable democracies do.

We can also quantify the magnitude of the effect of the predicted hazard on the changes in

∆FTA. With the quadratic term of the hazard, this effect is nonlinear. At the mean value of the

hazard (0.023 as shown in Table 1), doubling the hazard from 0.023 to 0.046 will lead to about one

percentage point increase in the FTA import share [i.e., (0.474-2*1.051*0.023)*0.023=0.01].

Column (3) of Table 3 reports the results from the analogous exercise for All-RTAs. Results

are largely unchanged, except that the second order hazard turns insignificant and the cutoff hazard

rate is lower (but still higher than more than 99% of the predicted hazards).

In the last two columns of Table 3A, we repeat the analysis in the previous two columns but

including DOM, FOR as well as duration dependence terms in the second stage RTA regressions.

The purpose of doing this is to check if our results are sensitive to the exclusion variables we used.

There we rely solely on the functional form (i.e. the nonlinearity of the first stage logit regression) to

identify the coefficients of the predicted hazard in the second stage. The results prove to be largely

insensitive to those changes. The excluded variables are highly insignificant except FOR.28

6.C. A bilateral analysis of the structural regression

We also run bilateral structural regressions. For this we use a large bilateral panel dataset

from Liu (2010), covering years 1960-2000. Each observation is associated with two countries as a

pair. A country pair appears only once in the data as single-dyad (either ij or ji). The dependent

variable is a dummy variable, which is equal to one if two countries had an RTA in a year and zero

otherwise. The RTA data are from the WTO’s RTA website and other sources (see footnotes in

Appendix 2). Our key explanatory variable is Hazard_avg, which is the average predicted hazard

27 These six extreme hazard rates are for Bangladesh and Burkina Faso. 28 We also tried excluding DOM and the duration dependence terms while including FOR in the second

stage; the results again hold well.

33

based on the enddemo logit regression in the first column of Table 3A.29 Other covariates include the

sum and the absolute difference of two countries’ GDP in a pair, the absolute difference in

GDP/capita, the great circle distance, land contingency or border dummy, the differences in labor

and environment standards (proxied by the absolute differences in their child labor ratios and

Carbon Dioxide/capita), common colony dummy, military conflict, political alliance, and the

product of the two countries’ remoteness as defined in equation (31).

The results from the bilateral analysis are reported in Table 3B. The first two columns are for

FTA/CU and All-RTA, respectively, using pooled data without country pair random effects; the last

two columns consider the unobserved heterogeneity by including country pair random effects. In all

the regressions, the average predicted hazard has a positive and significant impact on RTA

formation. Most of the other variables have the expected signs as well.30

6.D. Endogeneity

To address the endogeneity problem of FTA in the enddemo regressions, we adopt the two-

step control-function approach, as suggested by Petrin and Train (2010). The results are reported in

Table 4. Column (1) reports the first stage RTA import share regression. The explanatory variables

include all the covariates in the enddemo regression plus an instrument—remoteness, as defined in

equation (31). The remoteness variable is highly significant in the regression, suggesting that it is a

strong instrument.

In the second stage enddemo regression (the second column of Table 4), we include as an

additional variable the predicted residual from the first stage FTA regression. The predicted residual

(i.e., the control function) turns out to be highly insignificant in the enddemo regression, which

implies that endogeneity is not a problem. Higher order terms of the residual, if included, are also

insignificant. Adding this control function has little impact on our estimates. The results with the

control function in column (2) of Table 4 are very similar to those without control function in column

(3) of Table 2A.

Strictly speaking, the standard errors of the control function reported in the second column

of Table 4 are under-estimated because the control function is estimated. The correct standard errors

could be obtained by bootstrapping and would be larger than those reported in Table 4. But it is not

29 When one of the countries in a pair has missing hazard data, the non-missing hazard of the other country

is taken as Hazard_avg. 30 We did not include the quadratic term of the average hazard because it is always highly insignificant.

34

necessary to calculate the bootstrapped standard errors, since the control function is already highly

insignificant even with the downward biased standard errors.31

6.E. Robustness

We discuss in this section the robustness of our results in several ways. First, some studies

show that openness affects democracy (e.g., Mansfield, Milner, and Rosendorff, 2003; Mansfield,

Milner, and Pevehouse, 2008). To make sure that our FTA measure is not capturing just that, we

include an openness variable (defined as (imports+exports)/GDP) from the Penn World Table 6.3

and a WTO membership variable as additional regressors to our existing enddemo regressions. The

results are reported in Table 5. Results are similar to those from Table 2A, and the two newly added

variables are always highly insignificant.

Another issue is whether it is really the expectation of regime change (proxied by the

predicted hazard) that induces more FTAs. One alternative is that the change in regime itself causes

more FTAs. Another is that the regime itself has an effect (i.e. being democratic induces more FTAs).

Our estimated hazard, recent changes in political regime and current democracy status ought to be

correlated. To test whether it is indeed the expectation of regime change that matters, we include in

the ∆FTA regression both the predicted hazard and one of the following variables: (1) demp_pt, a

dummy indicating the current democracy status (i.e., whether polity>0); (2) reg_change, a dummy

indicating if a country’s polity scores changes signs (from strictly positive to non-positive or vice

versa); and (3) var(Polity)_10yr, the variance of polity scores during the last 10 years, which is used to

capture the regime stability in the previous 10 years. The results are reported in Table 6. In all the

three cases, the newly added variable is always highly insignificant, while the predicted hazard and

its quadratic term remain statistically significant. These results suggest that what really matters for

RTA formation is indeed the expectation of a regime change.

It is also possible that the empirical results may be actually identifying a more general

phenomenon than the model suggests. For example, it is possible that governments may form more

FTAs as a consequence of general political competition within an electoral system, and not

specifically because of the threat from autocrats. If the risk of democratic collapse is positively

associated with more political competition within a democracy, our finding that the threat of the

democratic system induces FTA formation may be driven by regular political competition within a

democracy. To test for that, we include a measure of political competition in the ∆FTA regression.

31 Under the null hypothesis that the control function is insignificant in the second stage, the t and F

statistics in the second stage remain valid.

35

We use the variable POLCOMP from Polity IV, which codes the degree of political competition in

the country.32 The results are reported in the last column of Table 6. Again, the newly added variable

is highly insignificant, while the predicted hazard and its quadratic term remain statistically

significant. This suggests that regular political competition in a country is not an important driver of

FTA formation, unlike the expectation of a regime change.

To further reinforce our confidence in the political economy explanations of our results, as a

falsification exercise, we run some regressions also for endauto (i.e. the end of autocracy spells)

instead of enddemo. The results are shown in Table 7. FTA variables are always highly insignificant.

As discussed above, given the tremendous heterogeneity in countries and FTAs, we believe

the FTA import share variable measures more precisely the importance of the FTAs than other

measures. It is nevertheless useful to check if our results still hold when alternative measures of

FTAs are used.33 We experiment first with a measure that counts the number of FTAs a country is

involved in a given year. Our findings hold very well. We also try a measure that counts the number

of FTA partners a country has in a given year, and a measure of the total economy size of all FTA

partners relative to the country’s own economy size. These last two measures do not yield as

consistent estimates as the other measures when the whole sample is used (estimates are often less

significant, although the signs of the coefficients are still expected in most cases), but do when the

poorest countries are excluded from our sample.34

Put together, these robustness checks boost confidence that we are capturing our intended

mechanism, rather than the workings of some omitted variable that affects both FTA formation and

political regimes in general.

7. CONCLUSIONS

[To be added]

32 This variable measures both the regulation of participation and the competitiveness in the political

process. According to the Polity IV Dataset User’s Manual, POLCOMP reaches its maximum score when “relatively stable and enduring political groups regularly compete for political influence and positions with little use of coercion; ruling groups and coalitions regularly, voluntarily transfer central power to competing groups; and no significant groups, issues, or types of conventional political action are regularly excluded from the political process.”

33 We do not report the results here to save space, but they are available from the authors. 34 Here the poorest countries are those with an average GDP/capita (in 2000 constant price) over 1960-2007

less than one dollar a day (similar to the poverty line defined by the World Bank).

36

REFERENCES

Aidt, Toke and Martin Gassebner (2008). “Do Autocratic States Trade Less?” Mimeo, University of Cambridge. Alesina, A. and G. Tabellini (1990). “A Positive Theory of Fiscal Deficits and Government Debt.” Review of Economic Studies 57, 403-414. Anderson, J., E. van Wincoop, 2003. “Gravity with gravitas: a solution of the border puzzle.” American Economic Review 93 (1), 170-192 Baier, Scott L., and Bergstrand, Jeffrey H. 2004. “Economic Determinants of Free Trade Agreements.” Journal of International Economics 64(1): 29-63. Banerji, A. and H. Ghanem (1997). “Does the Type of Political Regime Matter for Trade and Labor Market Policies?” World Bank Economic Review 11(1), 171-194. Besley, T. and S. Coate (2001). “Lobbying and welfare in a representative democracy,” Review of Economic Studies 68(1), 67-82. Dixit, Avinash (2008). “Democracy, Autocracy, and Bureaucracy.” Mimeo, Princeton University. Estevadeordal, Antoni, Caroline Freund, and Emanuel Ornelas (2008). "Does Regionalism Affect Trade Liberalization towards Non-Members?" Quarterly Journal of Economics 123(4), 1531-1575. Huntington, Samuel P. 1991. The Third Wave: Democratization in the Late Twentieth Century (Norman: University of Oklahoma Press). La Porta, Raphael., Lopez-De-Silanes, Franscico, Shleifer, Andrei, and Vishny, Robert (1999). “The Quality of Government”, Journal of Law, Economics and Organization 15, 222-79. Liu, Xuepeng, 2008. “The Political Economy of Free Trade Agreements: An Empirical Investigation,” Journal of Economic Integration 23(2), 237-271. Liu, Xuepeng, 2010. “Testing Conflicting Political Economy Theories: Full-fledged vs. Partial-scope Regional,” Southern Economic Journal 77(1): 78-103. Maddala, G.S. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge, MA: Cambridge University Press, 1983. Mansfield, E. D., H. V. Milner and J. C. Pevehouse (2008), “Democracy, Veto Players, and the Depth of Regional Integration.” World Economy 31(1) 67-96. 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. Mansfield, Edward D., Helen V. Milner, and B. Peter Rosendorff (2003). “Why Democracies Cooperate More: Electoral Control and International Trade Agreements.” International Organization 56(3): 477–513.

37

McLaren, John (2002). “A Theory of Insidious Regionalism.” Quarterly Journal of Economics 117(2), 571-608. Mitra, D., D. Thomakos and M. Ulubasogly (2002). “‘Protection for Sale’ in a Developing Country: Democracy vs. Dictatorship.” Review of Economics and Statistics 84(3), 497-508. Nalin, E. and J. Torstensson (1995). “Political Systems and Distortions: An Empirical Study.” Public Choice 84(1-2), 163-180. North, Douglass (1990). “A Transaction Cost Theory of Politics.” Journal of Theoretical Politics 2(4), 355-367. Olson, Mancur (1993). “Dictatorship, Democracy and Development.” American Political Science Review 87(3), 567-76. Ornelas, Emanuel (2005). "Rent Destruction and the Political Viability of Free Trade Agreements." Quarterly Journal of Economics 120(4), 1475-1506. Ornelas, Emanuel (2007). “Political Competition and the Strategic Adoption of Free Trade Agreements.” Mimeo, London School of Economics. Persson, T. and L. Svensson (1989). “Why a Stubborn Conservative Would Run a Deficit: Policy with Time-Inconsistent Preferences.” Quarterly Journal of Economics 104, 325-345. Persson, Torsten and Guido Tabellini (2009). “Democratic capital: The nexus of political and economic change.” American Economic Journal: Macroeconomics 1, 88-126. Petrin, Amil and Kenneth Train, 2010. "A Control Function Approach to Endogeneity in Consumer Choice Models,'' Journal of Marketing Research 47(1): 3-13. Pevehouse, Jon, 2002. “With a little help from my friends? Regional organizations and the consolidation of democracy.” American Journal of Political Science 46(3): 611-626. Rama, M. (1994). “Endogenous Trade Policy: A Time-Series Approach.” Economics and Politics 6(3), 215-231.

38

Figure 1: Number of RTAs in force and Cumulative Number of New Democracies (1948-2007)

Notes: See notes in Appendix 2 for the data sources of RTAs; New democracy is defined as an change from non-positive Polity score to a positive Polity score. Not all of these democracies spells are covered in our empirical analysis.

Figure 2: Nonparametric Survival Curves for Countries with and without FTA&CU

Notes: The curves are based on a sample with 2603 observations used in most of the regressions in Table 2. The curve on the top is the survival curve for countries with positive FTA&CU import shares (i.e., FTA_CU_Import_Share dummy=1; 1463 observations). The curve at the bottom is the survival curve for countries with zero FTA&CU import share (i.e., FTA_CU_Import_Share dummy=0; 1140 observations).

Num. of RTAs in Force (Right Y-axis)

Cumulative Number of New Democracies

(Left Y-axis)

0

20

40

60

80

100

120

1948

1950

1952

1954

1956

1958

1960

1962

1964

1966

1968

1970

1972

1974

1976

1978

1980

1982

1984

1986

1988

1990

1992

1994

1996

1998

2000

2002

2004

2006

0

50

100

150

200

250

0.00

0.25

0.50

0.75

1.00

0 20 40 60analysis time

FTA_CU_Import_Share = 0 FTA_CU_Import_Share = 1

Kaplan-Meier survival estimates, by FTA_CU_Import_Share

39

Figure 3: Histogram of the Predicted Hazard

Notes: The predicted hazard is based on the first regression in Table 3A.

Table 1: Descriptive statistics of variables

Variable Description Mean S.D. Min Max

enddemo Dummy indicating the end of a democracy 0.024 0.153 0 1

lFTA_impsh Lagged import share from FTA/CU partners 0.191 0.268 0 0.925

lRTA_impsh Lagged import share from All RTA partners 0.256 0.270 0 0.928

∆FTA_impsh Change in FTA_impsh from previous year 0.011 0.073 -0.451 0.772

∆RTA_impsh Change in RTA_impsh from previous year 0.013 0.074 -0.451 0.755

llgdppc Lagged log(GDP/capita) 7.830 1.471 4.400 10.540

ldc95_pt_cur Lagged current domestic democratic capital 0.535 0.338 0 0.99997

lforeigncap Lagged foreign democratic capital 0.045 0.138 -0.217 0.261

lwar Lagged war indicator 0.070 0.254 0 1

legor3 leg_origin=Socialism 0.120 0.325 0 1

africa Africa dummy 0.189 0.391 0 1

middleeast Middle East region dummy 0.033 0.180 0 1

socialist_trans Socialist transition dummy 0.101 0.301 0 1

Spain_colony UK colony dummy 0.184 0.388 0 1

UK_colony Spain colony dummy 0.344 0.475 0 1

duration Duration of a democracy (# of years passed) 20.211 15.967 1 60

Hazard Predicted hazard used in structural regressions 0.023 0.035 7.81e-11 0.379

Note: This table is based on a sample with 2603 observations used in most of the regressions in Table 2.

010

2030

4050

Den

sity

0 .1 .2 .3 .4Predicted Hazard

40

Table 2A: Enddemo regression results, import share measure (FTAs & CUs)

(1) (2) (3) (4) (5) (6)

logit logit xtlogit Weibull PH Cox xtreg (LPM)

lFTA_impsh -4.809*** -2.256* -2.697** -2.379** -2.155* -0.017** (1.187) (1.194) (1.306) (1.157) (1.103) (0.007) llgdppc -0.606*** -0.778*** -0.639*** -0.604*** -0.021 (0.157) (0.224) (0.159) (0.140) (0.013) ldc95_pt_cur -6.892** -9.323* -17.260*** -3.983** 0.059* (2.940) (5.544) (4.994) (1.746) (0.035) lforeigncap -5.031*** -5.871*** -5.815*** -4.869*** -0.075* (1.623) (1.636) (1.657) (1.426) (0.040) lwar 0.557 0.529 0.319 0.384 -0.011 (0.378) (0.481) (0.376) (0.371) (0.012) legor3 0.049 0.168 0.016 0.021 (0.888) (0.984) (0.863) (0.809) africa 0.248 0.340 0.206 0.166 (0.385) (0.510) (0.391) (0.344) middleeast 0.502 0.447 0.682 0.573 (1.035) (1.268) (0.953) (0.913) socialist_trans -0.056 -0.060 0.049 -0.046 (1.092) (1.119) (1.058) (0.999) Spain_colony -0.451 -0.391 -0.473 -0.431 (0.486) (0.536) (0.474) (0.448) UK_colony -1.184*** -1.414*** -1.268*** -1.168*** (0.415) (0.504) (0.436) (0.376) Durarion 0.456*** 0.657** 0.003*** (0.143) (0.285) (0.001) Duration^2 -0.010*** -0.013*** -0.000** (0.003) (0.005) (0.000) Cutoff Years 23 25 Country*Demo Spell RE Yes Country*Demo Spell FE Yes rho 0.205* 0.728 Test rho=0 (p-value) [0.086]

Shape parameter (φ) 5.601*** (1.306)

Pseudo R2 0.054 0.186 0.096 Log Lik -398.1 -238.5 -237.6 -104.3 -249.7 Observations 3,458 2,603 2,603 2,603 2,603 2,603

Notes: Standard errors in parentheses; Standard errors are clustered by democracy spells in logit regressions; Robust standard errors are used for the last three regressions; *** p<0.01, ** p<0.05, * p<0.1

41

Table 2B: Enddemo regression results, import share measure (All RTAs)

(1) (2) (3) (4) (5) (6)

logit logit xtlogit Weibull PH Cox xtreg (LPM)

lRTA_impsh -3.822*** -1.785* -2.058** -2.038** -1.663* -0.010 (0.762) (0.926) (1.026) (0.933) (0.865) (0.011) llgdppc -0.622*** -0.782*** -0.659*** -0.614*** -0.022 (0.159) (0.224) (0.162) (0.141) (0.013) ldc95_pt_cur -6.913** -9.332* -17.598*** -4.149** 0.062* (2.858) (5.548) (5.018) (1.736) (0.035) lforeigncap -4.501*** -5.206*** -5.139*** -4.394*** -0.081** (1.638) (1.701) (1.674) (1.427) (0.039) lwar 0.542 0.526 0.299 0.357 -0.011 (0.378) (0.476) (0.377) (0.372) (0.012) legor3 -0.164 -0.127 -0.245 -0.194 (0.880) (0.963) (0.856) (0.791) africa 0.155 0.206 0.085 0.085 (0.393) (0.507) (0.395) (0.348) middleeast 0.861 0.816 1.096 0.850 (1.018) (1.262) (0.890) (0.909) socialist_trans -0.112 -0.100 -0.007 -0.097 (1.055) (1.106) (1.023) (0.962) Spain_colony -0.390 -0.343 -0.416 -0.372 (0.485) (0.528) (0.477) (0.445) UK_colony -1.133*** -1.319*** -1.208*** -1.115*** (0.421) (0.499) (0.442) (0.380) Durarion 0.460*** 0.656** 0.003*** (0.142) (0.290) (0.001) Duration^2 -0.010*** -0.013*** -0.000** (0.003) (0.005) (0.000) Cutoff Years 23 25 Country*Demo Spell RE Yes Country*Demo Spell FE Yes rho 0.192 0.736 Test rho=0 (p-value) [0.115]

Shape parameter (φ) 5.992*** (1.338)

Pseudo R2 0.055 0.185 0.095 Log Lik -397.7 -238.7 -237.9 -104.0 -249.9 Observations 3,458 2,603 2,603 2,603 2,603 2,603

Notes: Standard errors in parentheses; Standard errors are clustered by democracy spells in logit regressions; Robust standard errors are used for the last three regressions; *** p<0.01, ** p<0.05, * p<0.1

42

Table 3A: Structural ∆RTA regression results

(1) (2) (3) (4) (5)

enddemo ∆FTA_impsh ∆RTA_impsh ∆FTA_impsh ∆RTA_impsh lFTA_impsh -0.145*** -0.156*** (0.021) (0.022) lRTA_impsh -0.161*** -0.166*** (0.022) (0.022) llgdppc -0.624*** 0.008 0.004 0.004 0.002 (0.153) (0.007) (0.007) (0.006) (0.007) ldc95_pt_cur -7.039** -0.019 0.004 (3.113) (0.022) (0.022) lforeigncap -5.550*** 0.240*** 0.180*** (1.601) (0.071) (0.066) lwar 0.533 -0.005 0.001 -0.005 0.000 (0.370) (0.006) (0.005) (0.005) (0.006) legor3 -0.131 (0.767) africa 0.300 (0.390) middleeast 0.682 (1.059) socialist_trans -0.350 (0.946) UK_colony -1.130*** (0.415) Spain_colony -0.329 (0.488) enddemo_pt_d1 0.452*** 0.001 -0.000 (0.151) (0.001) (0.001) enddemo_pt_d2 -0.010*** -0.000 0.000 (0.003) (0.000) (0.000) remote -0.697* -0.917** -1.151*** -1.323*** (0.397) (0.403) (0.415) (0.430) Hazard 0.474*** 0.407** 0.422** 0.417*** (0.156) (0.168) (0.192) (0.193) Hazard^2 -1.051* -1.330 -0.902 -1.293 (0.643) (0.849) (0.928) (0.837) Hazard Turning Point 0.225 0.153 0.234 0.161 Year Dummies Yes Yes Yes Yes Country FE Yes Yes Yes Yes R-squared 0.162 0.176 0.168 0.179 Observations 2,625 2,603 2,603 2,603 2,603

Notes: Standard errors in parentheses; Hazard is the predicted hazard based on the enddemo logit regression in the first column; Standard errors are clustered by country*spells in logit regression in column (1); Standard errors in the second stage (columns (2)-(5)) are corrected by bootstrapping (100 replications); *** p<0.01, ** p<0.05, * p<0.1

43

Table 3B: Bilateral Structural Regression Results

(1) (2) (3) (4)

FTA&CU (logit)

All RTAs (logit)

FTA&CU (xtlogit)

All RTAs (xtlogit)

Hazard_avg 5.447** 6.401*** 5.372** 6.123*** (2.159) (1.020) (2.646) (0.980) gdpsum 0.072*** 0.218*** 0.073*** 0.229*** (0.027) (0.013) (0.025) (0.015) dgdp -0.001 -0.173*** -0.002 -0.179*** (0.040) (0.024) (0.040) (0.025) dlgdppc_abs -0.100 -0.328*** -0.103 -0.346*** (0.076) (0.053) (0.091) (0.062) ldist -1.035*** -0.892*** -1.046*** -0.931*** (0.076) (0.046) (0.079) (0.053) border -0.262 0.192 -0.256 0.211 (0.218) (0.152) (0.203) (0.155) dchild -0.020** -0.007* -0.020** -0.007* (0.009) (0.004) (0.009) (0.004) dco2pc -0.516*** -0.154*** -0.520*** -0.153*** (0.078) (0.040) (0.083) (0.041) comcol 0.414** 0.091 0.426** 0.104 (0.203) (0.098) (0.189) (0.097) hostmean -1.794*** -1.057*** -1.822*** -1.075*** (0.610) (0.358) (0.604) (0.278) alliance 2.233*** 0.491*** 2.243*** 0.527*** (0.146) (0.099) (0.126) (0.094) remote -20.556*** 3.299*** -20.640*** 3.784*** (1.571) (0.686) (1.495) (0.733) Duration dependence

Yes Yes Yes Yes

Country Pair RE Yes Yes rho 0.016* 0.061* Test rho=0 [p-value] [0.086] [0.098] Pseudo R2 0.367 0.179 Log Lik -2130 -5588 -2131 -5587 Observations 186,502 176,591 186,502 176,591

Bilateral Data Source (1961-2000): Liu (2010) Notes: Standard errors in parentheses (not yet corrected by bootstrapping); *** p<0.01, ** p<0.05, * p<0.1; Hazard_avg is the average predicted hazard based on the enddemo logit regression in the first column of the previous table; When one of the country in a pair has missing hazard data, then the non-missing hazard of the other country is taken as Hazard_avg; Duration dependence is captured by a polynomial of the duration of an democracy up to six order.

44

Table 4: Endogeneity of RTA in Enddemo regressions (A control function approach)

(1) (2)

lFTA_impsh (xtreg)

enddemo_cf (xtlogit)

lFTA_impsh -2.830* (1.678) llgdppc 0.073*** -0.773*** (0.007) (0.227) ldc95_pt_cur 0.148*** -9.276* (0.019) (5.543) lforeigncap 0.498*** -5.788*** (0.093) (1.760) lwar 0.008 0.530 (0.007) (0.481) legor3 0.171 (0.984) africa 0.344 (0.511) middleeast 0.444 (1.269) socialist_trans -0.036 (1.136) UK_colony -1.426*** (0.514) Spain_colony -0.412 (0.561) Durarion 0.656** (0.285) Duration^2 -0.013*** (0.005) remote 3.771*** (0.655) Predicted Residual 0.246 (1.933) Country FE Yes Country*Demo Spell RE Yes R-squared 0.331 Observations 4,899 2,603

Notes: Standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1

45

Table 5: Robustness Check (1), Adding WTO Membership and Openness Variables

(1) (2) (3) (4)

xtlogit xtreg Weibull PH Cox lFTA_impsh -2.293* -0.017** -1.833* -1.799* (1.332) (0.007) (1.107) (1.092) llgdppc -0.728*** -0.016 -0.558*** -0.560*** (0.234) (0.013) (0.171) (0.150) ldc95_pt_cur -9.537* 0.055 -18.053*** -4.320** (5.517) (0.035) (4.983) (1.815) lforeigncap -5.372*** -0.064 -5.335*** -4.561*** (1.766) (0.041) (1.722) (1.523) lwar 0.314 -0.010 0.095 0.216 (0.529) (0.012) (0.394) (0.385) legor3 0.232 0.236 0.091 (1.005) (0.843) (0.809) africa 0.372 0.310 0.204 (0.522) (0.420) (0.360) middleeast 0.315 0.563 0.467 (1.306) (0.911) (0.887) socialist_trans -0.476 -0.513 -0.465 (1.205) (1.119) (1.086) Spain_colony -0.387 -0.500 -0.441 (0.560) (0.543) (0.485) UK_colony -1.237** -1.066** -1.043** (0.543) (0.472) (0.417) Lagged Openness -0.005 -0.000 -0.008 -0.004 (0.006) (0.000) (0.007) (0.007) WTO Membership -0.156 -0.010 -0.169 -0.130 (0.389) (0.008) (0.333) (0.280) Duration 0.676** 0.004*** (0.284) (0.001) Duration^2 -0.014*** -0.000** (0.005) (0.000) Cutoff Years 24 Country*Demo Spell RE Yes Country*Demo Spell FE Yes rho 0.211* 0.736 Test rho=0 [p-value] [0.081]

Shape parameter (φ) 5.866*** (1.323)

Pseudo R2 0.098 Log Lik -231.2 -99.94 -239.9 Observations 2,589 2,589 2,589 2,589

Notes: Standard errors in parentheses; Standard errors are clustered by democracy spells in logit regressions; Robust standard errors are used for LMP in column (5); *** p<0.01, ** p<0.05, * p<0.1

46

Table 6: Robustness Check (2), expectation for regime change vs. regime (change) itself, and political competition

(1) (2) (3) (4) (5)

enddemo ∆FTA_impsh ∆FTA_impsh ∆FTA_impsh ∆FTA_impsh lFTA_impsh -0.145*** -0.145*** -0.146*** -0.149*** (0.021) (0.021) (0.021) (0.021) llgdppc -0.624*** 0.008 0.008 0.008 0.008 (0.153) (0.007) (0.007) (0.007) (0.007) ldc95_pt_cur -7.039** (3.113) lforeigncap -5.550*** (1.601) lwar 0.533 -0.005 -0.005 -0.005 -0.005 (0.370) (0.005) (0.005) (0.005) (0.005) legor3 -0.131 (0.767) africa 0.300 (0.390) middleeast 0.682 (1.059) socialist_trans -0.350 (0.946) UK_colony -1.130*** (0.415) Spain_colony -0.329 (0.488) Durarion 0.452*** (0.151) Duration^2 -0.010*** (0.003) remote -0.695* -0.708* -0.722* -0.734* (0.379) (0.380) (0.384) (0.390) Hazard 0.473*** 0.499*** 0.504*** 0.524*** (0.117) (0.123) (0.128) (0.125) Hazard^2 -1.058** -1.141*** -1.163** -1.304** (0.429) (0.437) (0.467) (0.522) demo_pt -0.002 (0.010) reg_change 0.007 (0.008) var(Polity)_10yr -0.000 (0.000) polcomp 0.003 (0.002)

Observations 2,625 2,603 2,603 2,603 2,557 R-squared 0.162 0.162 0.162 0.167

Notes: Standard errors in parentheses (not yet corrected by bootstrapping); Standard errors are clustered by democracy spells in logit regression in column (1); *** p<0.01, ** p<0.05, * p<0.1; demp_pt is a dummy indicating the current democracy status (i.e., polity>0); reg_change is a dummy indicating if a country’s polity score changes sign (from positive to negative or vice versa); var(Polity)_10yr is the variance of polity scores during the last 10 years; polcomp measures the degree of political competition in a country.

47

Table 7: Robustness Check (3), End of Autocracy (endauto)

(1) (2) (3) (4)

xtlogit Weibull PH Cox xtreg (LPM) lFTA_impsh -1.834 -0.727 -0.673 -0.016 (2.072) (0.762) (0.763) (0.014) llgdppc -0.839** -0.139 -0.129 -0.046*** (0.380) (0.109) (0.113) (0.009) ldc95_pt_cur 1.057 3.463*** 2.789*** -0.526** (2.527) (1.081) (0.962) (0.230) lforeigncap -1.027 3.296*** 2.994*** -0.313*** (2.087) (0.936) (1.091) (0.060) lwar 0.163 0.488 0.442 0.001 (0.656) (0.323) (0.322) (0.011) legor3 -1.921 -0.183 -0.216 (1.736) (0.595) (0.584) africa -3.253*** -1.241*** -1.220*** (1.079) (0.324) (0.324) middleeast -7.546*** -2.541*** -2.535*** (2.174) (0.937) (0.897) socialist_trans 0.506 -0.043 -0.014 (2.017) (0.650) (0.629) Spain_colony 0.297 -0.176 -0.098 (1.215) (0.365) (0.376) UK_colony 0.567 0.169 0.292 (0.981) (0.318) (0.316) Auto_Duation 0.176*** 0.002* (0.059) (0.001) Auto_Duation^2 -0.001 0.000* (0.001) (0.000) Country*Auto Spell RE Yes Country*Auto Spell FE Yes Pseudo R2 0.0621 Log Lik -254.7 -129.8 -281.2 Observations 2,435 2,253 2,253 2,435

Notes: Standard errors in parentheses; Robust standard errors are used for the last three regressions; *** p<0.01, ** p<0.05, * p<0.1

48

Appendix 1: Countries covered in the analysis (Total = 133)

Albania France New Zealand

Algeria Gambia Nicaragua

Angola Georgia Niger

Argentina Germany Nigeria

Armenia Ghana Norway

Australia Greece Panama

Austria Guatemala Papua New Guinea

Azerbaijan Guinea-Bissau Paraguay

Bangladesh Guyana Peru

Belarus Haiti Philippines

Belgium Honduras Poland

Benin Hungary Portugal

Bolivia India Qatar

Botswana Indonesia Romania

Brazil Iran Russia

Bulgaria Ireland Senegal

Burkina Faso Israel Sierra Leone

Burma Italy Singapore

Burundi Jamaica Slovakia

Cambodia Japan Slovenia

Cameroon Kenya Somalia

Canada Korea, South South Africa

Central African Rep. Kuwait Spain

Chile Kyrgyzstan Sri Lanka

Colombia Laos Sudan

Comoros Latvia Sweden

Congo, Dem. of Lebanon Switzerland

Congo, Rep. of Lesotho Syria

Costa Rica Liberia Taiwan

Cote D Ivoire Lithuania Tanzania

Croatia Macedonia Thailand

Cuba Madagascar Trinidad and Tobago

Cyprus Malawi Tunisia

Czech Rep. Malaysia Turkey

Czechoslovakia Mali Uganda

Denmark Mauritania Ukraine

Djibouti Mauritius United Arab Emirates

Dominican Rep. Mexico UK

Ecuador Moldova USA

Egypt Mongolia Uruguay

El Salvador Morocco Venezuela

Equatorial Guinea Mozambique Zambia

Estonia Namibia Zimbabwe

Fiji Nepal

Finland Netherlands

Note: This table lists the countries covered by the first regression in Table 2A.

49

Appendix 2: List of RTAs covered in this paper (1948-2007; Total = 219)

RTA Names RTA Type

RTA Names RTA Type

RTA Names RTA Type

ACM CU EC-Egypt FTA Kyrgyz-Armenia FTA

AFTA FTA EC-Estonia FTA Kyrgyz-Kazakhstan FTA

AMU FTA EC-FYROM FTA Kyrgyz-Moldova FTA

Albania-Moldova FTA EC-Hungary FTA Kyrgyz-Russia FTA

Albania-Bosnia & Herzeg. FTA EC-Iceland FTA Kyrgyz-Ukraine FTA

Albania-FYROM FTA EC-Israel FTA Kyrgyz-Uzbekistan FTA

Albania- Serbia & Mont. FTA EC-Jordan FTA LAIA PTA

Arab Free Trade Area FTA EC-Latvia FTA Laos-Thailand PTA

Armenia-Kazakhstan FTA EC-Lebanon FTA MERCOSUR FTA

Armenia-Moldova FTA EC-Lithuania FTA MRU CU

Armenia-Russia FTA EC-Malta FTA MSG PTA

Armenia- Turkmenistan FTA EC-Mexico FTA Macedonia-Bosnia & Herzeg. FTA

Armenia- Ukraine FTA EC-Morocco FTA Mexico-Nicaragua FTA

BAFTA FTA EC-Norway FTA Mexico-Costa Rica FTA

Bangkok PTA EC-Poland FTA Mexico-Israel FTA

Bangkok-China PTA EC-Romania FTA Moldova-Croatia FTA

Bulgaria-Albania FTA EC-Slovak FTA Moldova-FYROM FTA

Bulgaria-Estonia FTA EC-Slovenia FTA Moldova-Serbia & Mont. FTA

Bulgaria-Israel FTA EC-South Africa FTA Moldova- Bosnia & Herzeg. FTA

Bulgaria-Latvia FTA EC-Switzerland FTA Morocco-Turkey FTA

Bulgaria-Lithuania FTA EC-Syria FTA NAFTA FTA

Bulgaria-Macedonia FTA EC-Turkey CU New Zealand-Singapore FTA

Bulgaria-Turkey FTA EC-Tunisia FTA PATCRA FTA

CACM FTA ECCAS FTA PTN PTA

CACM-Chile FTA ECO PTA Pan-Arab Free Trade FTA

CACM-Costa Rica FTA ECOWAS FTA Panama-El Salvador FTA

CAFTA-Dominican Rep. FTA EFTA FTA Panama-Singapore FTA

CAN FTA EFTA-Chile FTA Panama-Taiwan FTA

CARICOM FTA EFTA-South Korea FTA Poland-Faroe Islands FTA

CARICOM-Bahamas FTA EFTA-Tunisia FTA Poland-Israel FTA

CARICOM-Colombia PTA EFTA-Bulgaria FTA Poland-Latvia FTA

CARICOM-Costa Rica FTA EFTA-Croatia FTA Poland-Lithuania FTA

CEFTA FTA EFTA-Czech FTA Poland-Turkey FTA

CEPGL FTA EFTA-Estonia FTA Rep. of Korea-Singapore FTA

CER FTA EFTA-Finland FTA Romania-Moldova FTA

CIS FTA EFTA-Hungary FTA SACU FTA

COMESA PTA EFTA-Israel FTA SADC FTA

Canada-Chile FTA EFTA-Jordan FTA SAPTA PTA

Canada-Costa Rica FTA EFTA-Latvia FTA SPARTECA PTA

Canada-Israel FTA EFTA-Lithuania FTA Singapore-Australia FTA

Ghana-Burkina Faso FTA EFTA-Macedonia FTA Slovak-Estonia FTA

Chile-China FTA EFTA-Mexico FTA Slovak-Israel FTA

Chile-Costa Rica FTA EFTA-Morocco FTA Slovak-Latvia FTA

Chile-El Salvador FTA EFTA-Poland FTA Slovak-Lithuania FTA

Chile-Korea FTA EFTA-Romania FTA Slovak-Turkey FTA

Chile-Mexico FTA EFTA-Singapore FTA Slovenia-Bosnia & Herzeg. FTA

50

RTA Names RTA Type

RTA Names RTA Type

RTA Names RTA Type

China-HK FTA EFTA-Slovak FTA Slovenia-Croatia FTA

China-Macao FTA EFTA-Slovenia FTA Slovenia-Estonia FTA

Croatia-Macedonia FTA EFTA-Turkey FTA Slovenia-Israel FTA

Croatia-Serbia & Mont. FTA El Salvador-Mexico FTA Slovenia-Latvia FTA

Croatia- Bosnia & Herzeg. FTA Estonia-Faroe Islands FTA Slovenia-Lithuania FTA

Croatia-Albania FTA Estonia-Turkey FTA Slovenia-Macedonia FTA

Czech-Estonia FTA Estonia-Ukraine FTA TRIPARTITE PTA

Czech-Israel FTA GCC PTA Thailand-Australia FTA

Czech-Latvia FTA GSTP PTA Thailand-New Zealand FTA

Czech-Lithuania FTA Georgia-Armenia FTA Trans-Pacific SEP FTA

Czech-Slovak CU Georgia-Azerbaijan FTA Turkey-Tunisia FTA

Czech-Turkey FTA Georgia-Kazakhstan FTA Turkey-Bosnia & Herzeg. FTA

Dominica-Costa Rica FTA Georgia-Russia FTA Turkey-Croatia FTA

Dominica-El Salvador FTA Georgia-Turkmenistan FTA Turkey-Israel FTA

Dominica-Guatemala FTA Georgia-Ukraine FTA Turkey-Latvia FTA

Dominica-Honduras FTA Guatemala-Mexico FTA Turkey-Lithuania FTA

EAC PTA Honduras-Mexico FTA Turkey-Macedonia FTA

EAEC CU Hungary-Estonia FTA Turkey-Slovenia FTA

EC CU Hungary-Israel FTA US-Canada FTA

EC-Albania FTA Hungary-Latvia FTA US-Chile FTA

EC-Algeria FTA Hungary-Lithuania FTA US-Israel FTA

EC-Andorra CU Hungary-Turkey FTA US-Jordan FTA

EC-Bulgaria FTA India-Sri Lanka FTA US-Singapore FTA

EC-Chile FTA Japan-Malaysia FTA United States-Australia FTA

EC-Croatia FTA Japan-Mexico FTA United States-Bahrain FTA

EC-Cyprus PTA Japan-Singapore FTA United States, Morocco FTA

EC-Czech FTA Jordan-Singapore FTA WAEMU PTA

Data Sources: (1). WTO: http://www.wto.org/english/tratop_e/region_e/region_e.htm (dated May 1, 2004 and July 18, 2007); (2). WTO Archive, WTO, Geneva, Switzerland; (3). Frankel, Stein and Wei (1997); (4). Schiff and Winters (2003); (5). Foreign Trade Information System: http://www.sice.oas.org/agreements_e.asp (accessed in 2004)

Notes: FTA – Free trade areas; CU- Customs union; PTA – Partial-scope RTAs; Accession agreements to existing agreements are not displayed separately; Service agreements are not counted separately; An agreements into force before July 15 of a year (inclusive) is considered as effective in this year, and considered as the following year otherwise.

51

Appendix 3: Cases of Enddemo Covered in this Paper (Total = 91)

Country Name Year Country Name Year Country Name Year

Albania 1996 Egypt 1952 Nigeria 1984

Algeria 1962 Equatorial Guinea 1969 Panama 1968

Angola 1975 Fiji 1987 Peru 1962

Argentina 1976 Fiji 2006 Peru 1968

Armenia 1996 Gambia 1994 Peru 1992

Azerbaijan 1993 Ghana 1960 Philippines 1972

Bangladesh 1974 Ghana 1972 Qatar 1971

Bangladesh 2007 Ghana 1981 Sierra Leone 1967

Belarus 1995 Greece 1967 Sierra Leone 1971

Benin 1963 Guatemala 1954 Sierra Leone 1997

Brazil 1964 Guatemala 1974 Somalia 1969

Burkina Faso 1980 Guinea-Bissau 1998 Sudan 1958

Burma 1962 Guinea-Bissau 2003 Sudan 1970

Cambodia 1997 Guyana 1978 Sudan 1989

Cameroon 1960 Haiti 1991 Syria 1949

Central African Rep. 2003 Haiti 2000 Syria 1951

Chile 1973 Indonesia 1950 Thailand 1971

Comoros 1976 Iran 2004 Thailand 1976

Comoros 1995 Kenya 1966 Thailand 1991

Comoros 1999 Korea, South 1961 Thailand 2006

Congo, Dem. of 1960 Korea, South 1972 Tunisia 1959

Congo, Rep. of 1963 Kuwait 1963 Turkey 1971

Congo, Rep. of 1997 Laos 1960 Turkey 1980

Cote D Ivoire 1960 Lebanon 1975 Uganda 1966

Cote D Ivoire 2002 Lesotho 1970 Uganda 1985

Cuba 1952 Lesotho 1998 United Arab Emirates 1971

Cyprus 1963 Madagascar 1960 Uruguay 1972

Djibouti 1977 Morocco 1956 Zambia 1968

Dominican Rep. 1963 Nepal 2002 Zimbabwe 1987

Ecuador 1961 Niger 1996

Ecuador 1970 Nigeria 1966

Notes: This table lists the enddemo covered by the first regression in Table 2; Enddemo refers to the end of a democracy (i.e., a change from a positive Polity score to a non-positive Policy score); Democracies includes both new and old democracies (as in Person and Tabellini’s (2008) definition).