Appendix: The Politics of Trade Agreement Design: The Obligation-Flexibility Trade-Off Leonardo Baccini, Princeton University Andreas D¨ ur, University of Salzburg Manfred Elsig, University of Bern
Factor analysis (ObligationII ) We implemented factor analysis as follows. First, we converted the seven categorical variables that we included in the operationalization of ObligationII into dummies.1 Second, we extracted two factors from the items and used oblique rotation to ensure high factor loadings.2 The choice of extracting two factors is ultimately discretionary. However, the rule of thumb is to check the plot of score variables to detect the number of factors (Costello and Osborne, 2005). Figure A-1 shows that two factors appear to be a sensible choice. Third, we use the first of the two factors that we extract as a measure of ObligationII, as it is highly correlated with a simple additive score across the variables that we used in the factor analysis (CoarseObligationII).3 Finally, the factor scores that we use as measure of ObligationII are calculated us1 For instance, the categorical variable Investment Coverage was converted into four dummies: general statement on investment protection (one if present), investment protection based on BIT (one if present), investment protection in the services chapter (one if present), and separate chapter for investment (one if present). 2 We drop those items that have a low correlation, i.e. between −0.3 and +0.3. By using oblique rotation we assume that the angle between the two factors is not rectangular. 3 We obtain similar results if we exclude five items with a high value (that is, higher than 0.7) on “uniqueness” in the factor loadings table (Kim and Mueller, 1978).
1
Figure A-1: Plot of the score variables with two factors. ing the regression method. Put simply, we multiply each factor loading with its respective item and we take the sum of all these products. We label this variable ObligationII.
To assess whether our operationalization of ObligationII make sense, we compare the measure of ObligationII that we obtain using factor analysis with the variable CoarseObligationII. Figure A-2 shows that the correlation is quite high (ρ = 0.9), confirming that our operationalization is a refinement of CoarseObligationII without being a completely different variable. Next to a graph with all the PTAs (on 2
Figure A-2: ObligationII with factor analysis versus CoarseObligationII. the right side) we show a graph (on the left side) with only those PTAs that have high values for both indicators of ObligationII and CoarseObligationII. This should simplify the reading of the graph. Moreover, the correlation between ObligationI and ObligationII is 0.5. Thus, although these two variables are (not surprisingly) highly correlated, they seem to capture two different dimensions of the design of PTAs. Figure A-3 shows this relationship for PTAs formed between 1990 and 1994 and between 2005 and 2009.
3
Figure A-3: ObligationII with factor analysis versus ObligationI for PTAs formed between 1990-94 (right side) and between 2005-09 (left side).
4
Control variables and descriptive statistics We use GDP per capita (GDP pc) and total GDP (GDP ) to capture the income level and economic importance of a country. Larger and richer countries might be able to negotiate agreements with a higher degree of obligation and less flexibility compared to poorer or smaller countries. Moreover, we include economic growth (GDP Growth), a variable that may again have an effect on both the degree of obligation and flexibility. As to obligation, Mattli (1999) argues that countries that experience an economic downturn may be more likely to accept a loss of sovereignty. Koremenos (2005) submits that countries that experience low economic growth are supposed to be more risk-acceptant than countries that experience an economic upturn and thus may be more willing to accept a rigid agreement.4 Moreover, we add T rade, which is the log of the value of exports between the two countries in the dyad. We expect the amount of trade to positively influence the demand for agreements with both a high degree of obligation and flexibility.5
We also include a democracy score (Regime), which comes from Cheibub et al. (2010).6 . Our expectation is that democracies should sign deeper and more rigid agreements than autocracies. Moreover, we add a variable that captures if a country has undergone a transition from autocracy to democracy (Democratization). This 4 Koremenos (2005) uses a different operationalization of risk-aversion, distinguishing between mid-growth and low-high growth. As robustness check, we also try her specification obtaining similar results. 5 The correlation between T rade and IIT is .4. 6 Results do not change if we use other measures such as Freedom House and Polity IV.
5
variable scores one if a country has become a democracy over the past five years using Polity IV as indicator. The sign of this variable is difficult to predict. On the one hand, democratizing countries might bargain more rigid agreements to enhance their credibility in the international system. On the other hand, democratizing countries might require more flexibility since they face high levels of uncertainty about future states of the world. Furthermore, we include the number of V etoP layers from Henisz (2000). As the number of veto players increases, so does the opposition to PTAs (Mansfield et al., 2007; Peterson and Thies, 2011). Thus, countries are expected to bargain PTAs with a low degree of obligation and a high degree of flexibility.
Furthermore, we include a measure of geographic distance (Distance), for two reasons. On the one hand, distance captures the commercial and strategic salience of a country for the other country in the dyad. On the other hand, monitoring may be easier for countries that are close to another compared to countries that are far away. We also add a dummy that scores one if a country is a WTO member (W T O). WTO members tend to implement trade policies that differ from countries that are not part of this international organization (Mansfield and Reinhardt, 2003). They also have ad hoc flexibility provisions upon which they can rely. Moreover, we include the degree of obligation and flexibility of an agreement (if any) that was previously negotiated by the dyad. The number of member countries of a PTA (N o.M embers) is another control variable as there may be a broader-deeper tradeoff in international agreements (Gilligan, 2004). 6
Table A-1 summarizes the descriptive statistics of the dependent and independent variables. Table A-1: Descriptive statistics. Variables Flexibility Tariff Flexibility ObligationI ObligationII PreviousFlex PreviousTariffFlex PreviousObligationI PreviousObligationII FlexIDiffusion TariffFlexDiffusion ObligationIIDiffusion ObligationIDiffusion USDesign Crisis IGOMembership IIT GDPpc GDP GDP Growth Trade Democracy Democrat. VetoPlayers WTO Distance No. Members
Mean .70 3.12 2.19 .52 .15 1.05 2.38 .61 .15 2.76 .30 .07 18.49 .97 3.14 .10 1.27 1.43 .97 2.16 .23 .10 .14 .54 7.99 40.7
Std. Dev. .33 1.09 2.86 .33 .34 1.61 .84 .32 .34 .74 1.38 .17 .89 .66 1.12 .14 2.71 1.12 3.61 2.40 .42 .30 .17 .70 .99 30.8
Min 0 0 0 0 0 0 .94 .26 0 1.16 0 0 17.13 0 0 0 .09 .20 -8.9 0 0 0 0 .46 4.47 2
Max 1 3.93 16.75 1 1 3.93 15.1 1 1 3.63 16.56 1 19.08 5 9 .70 35.62 6.50 9.5 12.25 1 1 .65 1 9.87 91
7
Instruments to Identify the Models To correctly identify our models we include several instruments. Three instruments are build upon diffusion variables. For F lexibility we also include a variable that capture how frequently a country faces a financial, banking or currency crisis. For Tariff F lexibility we include a variable that captures the number of joint membership in economic (but not trade) IGOs between two countries in a dyad. Below we explain these instruments and their strengths in detail.
Diffusion Instruments First, we include variables that capture the diffusion of the PTA design. In designing PTA treaties countries are likely to be affected by PTAs signed by competitors. For instance, a Colombian negotiator, interviewed by the authors in February 2012, argues that the design of the Colombia-US PTAs was influenced by the design of the Chile-US PTA and NAFTA. Similarly, South Africa insisted for bargaining with the EU a PTA in line the ones formed by some North African countries after the Barcelona Euro-Mediterranean Ministerial Conference (1995).7
We operationalize competition as a function of PTA type (for which we distinguish three types, namely bilateral, plurilateral among countries from the same region and plurilateral among countries from different regions) and level of develop7
Confirmed by interviews with several South Africa negotiators and policy-makers, January and February 2012.
8
ment (namely, North-North, North-South and South-South). Concretely, these variables capture the idea that in reacting to PTAs countries care especially about agreements signed by competitors (usually neighboring countries) and about far-reaching agreements (with north-south PTAs being usually more far-reaching than southsouth PTAs). More generally, the level of obligation and flexibility competition variables (labeled as ObligationDiffusion, F lexibilityDiffusion and Tariff F lexDiffusion) control for the fact that the formation and design of a PTA have an impact on the formation and design of other PTAs (Baccini and D¨ ur, 2012).
Since each of these variables appear in only one of the two equations (e.g. ObligationIIDiffusion in ObligationII), they allow for the model to be identified. Each of these instruments is a strong predictor of the respective dependent variable. Specifically, the correlation between F lexibility and F lexibilityDiffusion is 0.67, the correlation between Tariff F lexibility and Tariff F lexDiffusion is 0.75, the correlation between ObligationII and ObligationIIDiffusion is 0.53, and the correlation between ObligationI and ObligationIDiffusion is 0.90. Conversely, each of these instruments is weakly correlated with the other two dependent variables in which it is not included. For instance, correlation between ObligationII and Tariff F lexDiffusion is 0. Finally, each of these instruments is weakly correlated with the error term of the other two dependent variables. For instance, ObligationII and Tariff F lexDiffusion is 0.1.
That said, we cannot completely rule out the possibility that any of these in9
struments is correlated with another dependent variable because our paper makes an argument about the endogenous design of PTAs. However, we contend that the causal chain would be so indirect that we can be confident that the exclusion restriction holds across our models. Moreover, interviews with policy-makers and negotiators highlight that in bargaining the level of obligations of a PTA countries are often concerned about the level of obligations of other PTAs formed by competitors and/or neighboring countries, but rarely take into account other dimensions of these competing PTAs, e.g. their degree of flexibility.8
Finally, we include a variable that captures the maximum value of the variable ObligationII for PTAs signed by the US. We label this variable U SDesign. There is indeed evidence that US PTAs are the deepest PTAs and that other countries include provisions borrowed by the US treaties. For instance, the EU changed quite significantly the design of its PTAs after the formation of NAFTA. In this regard, the EU-Mexico trade agreement is usually defined a NAFTA-like PTA (Zapanta, 2000).9 Other countries like Japan also followed the US in design their PTAs (Manger, 2005).
The variable U SDesign is high correlated with ObligationII and weakly correlated with the other design variables. We note that for ObligationI and F lexibility there is no variation in the US PTAs, i.e. they all score the highest value over the period under investigation. Moreover, we do not include a variable that captures 8 9
Interviews carried out by the authors in December 2011 and in January 2012. Document available at http://www.usmcoc.org/b-nafta5.php.
10
the maximum value if the variable Tariff F lexibility. It would make sense to include such a variable only if data on tariff transition were available at product level. However, since data are at aggregate level, identifying the effect of tariff transition of US PTAs on tariff transition of PTAs signed by other countries is very problematic.10
Flexibility: Crisis Flexibility concerns provisions that allow a country to stop cooperating if it faces “exceptional circumstance”. As such, countries that face frequent financial crisis are expected to be more risk adverse and to design agreements that include high levels of flexibility (Koremenos, 2005). We include a count variable that calculate the total number of banking, currency, and financial crisis faced by a country. The variable has a maximum of five, scored by Brazil and Argentina. We take the maximum of the two values in a dyad. Data come from Laeven and Valencia (2008).
We note that there are no theoretical reasons to believe that the frequency of crisis should affect the level of obligations in a PTA. Indeed, obligations included into a treaty are meant to work in normal times. Moreover, the variable GDP Growth included in model in which ObligationI and ObligationII are dependent variable should control for any contingent effect of economic conditions on cooperation.
Moreover, Tariff F lexibility should not be affect by the the variable Crisis, 10
If we include such a variable, results do not change and the instrument is not statistically significant.
11
since Tariff F lexibility is not a temporary trade barrier to deal with “exceptional circumstance”, but rather a device to spread adjustment costs coming from trade liberalization over a longer period of time. Our hunches are confirmed by the fact that Crisis is highly correlated with F lexibility and weakly correlated with the other design variables.
Tariff Flexibility: Joint Membership in Economic IGOs We include a variable that count the number of joint membership in economic IGOs, which do not regulate the trade sector. We label this variable IGOM embership. Data on IGOs come from Pevehouse et al. (2004), whereas data on IGOs categorization come from Ingram et al. (2005). Countries that share a large number of memberships are used to cooperate one another. Such an historical record of cooperation makes easier the enforcement of a further agreement or, at least, improves the perception that the agreement will be fulfilled by the trade partners. In turn, as trust in the trade partner increases due to a positive historical record of cooperation, countries are more likely to grant each other a longer transition period before of the full implementation of the PTA. As such, this variable is expected to be a good predictor of Tariff F lexibility.
We note that this the IGOM embership variable captures the level of trust between countries based on previous cooperation, but does not capture how easy monitoring an agreement is (which would also be theoretically related to the level
12
of obligations). Moreover, the IGOM embership variable does not include trade organizations. Thus, after having controlled for the presence of a previous PTA between two countries, the IGOM embership variable should not effect ObligationI and ObligationII. Indeed, IGOM embership is highly correlated with F lexibility and weakly correlated with the other design variables.
13
Additional evidence: Flexibility/Obligation Trade-Off Table below shows the Flexibility/Obligation trade-off faced by countries when they design PTA. Table A-2: PTAs among countries F lexibility/ObligationII trade-off. Low Flex & Low ObligationII Central American Integration System† Central European Free Trade Agreement† Common Economic Zone† Economic and Monetary Community of Central Africa† Eurasian Economic Community† Gulf Cooperation Council† Argentina-Uruguay Armenia-Georgia Armenia-Iran Bolivia-Uruguay Brazil-Suriname Bhutan-Uruguay Bhutan-India Estonia-Norway Georgia-Kazakhstan Guatemala-Mexico India-Nepal Israel-Jordan Jordan-Lebanon Jordan-Libya Kazakhstan-Kyrgyzstan Kyrgyzstan-Russia Laos-Thailand Paraguay-Venezuela Turkmenistan-Ukraine Uruguay-Venezuela
14
with
high
level
High Flex & High ObligationII Bulgaria-EU Czech Rep.-EU Hungary-EU Lithuania-EU Poland-EU Slovakia-EU Bolivia-Chile Czech Rep.-Slovenia Latvia-Poland Latvia-Slovakia
of
IIT:
the
Matching Matching requires a series of discretionary decisions. First, which variables should be selected to match between the treatment group and the control group? We follow two criteria. We select variables that are (1) important drivers of PTA formation and (2) theoretically correlated with flexibility, ObligationI and ObligationII. To begin with, we match on IIT for the reasons explained in the main text. Moreover, we match on Distance, T rade, GDP pc, and BIT . Previous studies show that these variables are crucial predictors of the formation of PTAs (Baier and Bergstrand, 2004; Baccini and D¨ ur, 2012) and logically related to the design of an agreement.
In addition to these economic variables, we match on two important political variables. First, we use the variable Regime because previous research indicates that democratic countries have stronger incentives to engage in trade cooperation with other democracies (Mansfield et al., 2002) and to design more rigid PTAs (Kucik, 2011). Second, we match on BIT , a variable that scores one if two countries are member of the same BIT. Indeed, Baccini and D¨ ur (2012) show that countries that formed a BIT are more likely to sign a PTA that includes strict regulations on investment and investment-related sectors, i.e. a deeper PTA. Third, we match on WTO membership because WTO members have already previously liberalized their trade policies, so they can form PTAs at a lower cost. Since the WTO includes flexibility clauses and provisions on investments, services, and intellectual property rights, WTO members may also have a greater propensity to include such provisions
15
into PTAs.
The second discretionary decision is: which coarsening should be chosen for these variables? To choose the coarsening, we examined the distribution of these variables. IIT , GDP pc, and T rade are skewed on the right, i.e. there are few rich big countries. Thus, we chose to coarsen these variables at their mean and one standard deviation above the mean. Conversely, Distance is skewed to the left, i.e. there are few countries very close to another. Thus, we coarse this variable at its mean and one standard deviation below the mean. The main advantage of doing so is that we can place outliers, i.e. rich and big countries, in the same bin. The Regime, BIT , and W T O variables cannot be coarsened since they are dummies. Figure A-4 shows the distribution of the continuous variables and the values at which coarsening has been chosen. Figure A-4 About Here Third, we identify those observations that contain at least one treated and one control unit and we drop all the others. To do so, we match our original number of dyads with the dyads that did not form a PTA. These dyads without PTAs come from all the possible dyadic combinations of the original 156 trading entities, i.e. 156×155 2
= 12, 090. Owing to matching, we lose 166 dyads and 57 PTAs.11 For in-
stance, we lose some bilateral agreements signed by the EU and the European Free Trade Association with East European countries. Among plurilateral trade agree11
The number of dyads that we lose doubles when we use directed dyads.
16
ments, we lose a few dyads from the Caribbean Community (CARICOM),the Central American Free Trade Agreement (CAFTA), the Central European Free Trade Agreement (CEFTA), and the Gulf Cooperation Council (GCC).
Figure A-5 shows the reduction in imbalance for each covariate looking both at the difference between the means and the ratio of the variances. As the figure shows, the matching dramatically improves this balance, and thus enhances our ability to identify effects in the data. More importantly, the overall L1 statistic measure, which captures imbalance with respect to the full joint distribution, drops significantly from 0.87 to 0.70. Both tests confirm the validity of our matching strategy in balancing our treated sub-sample with our control sub-sample. Figure A-5 About Here
17
Figure A-4: Coarsening of continuous variables for the matching analysis.
18
Figure A-5: The balance of mean and variance between treated and untreated observations in the data, before and after matching.
19