FTA and Economic Growth: A Nonparametric Approach

Transcription

1 FTA and Economic Growth: A Nonparametric Approach Jung Hur * Sogang University Cheolbeom Park ** Korea University April 14, 2009 Abstract This paper examines whether a bilateral FTA exerts a positive effect on the economic growth of the two countries involved in the FTA. It employs a nonparametric matching approach, which imposes no specific functional forms on the relation, and can be applied to a broad range of data structures. Unlike the results from earlier linear regression model approaches, we find an insignificant effect of FTA on the total economic growth of the FTA. In particular, we show that an FTA exerts no significant effects from a one-year to ten-year horizon after its launch. Furthermore, we detect an upward trend of a gap between the growth rates of per capita real GDP among countries within an FTA. This implies uneven FTA effects across countries within an FTA, which may explain, in part, the insignificant effect of an FTA on the total economic growth of an FTA. Key Words: Free trade agreement, Growth, Matching, Treatment effect JEL Classification: F13, O49, C14, C21 * Department of Economics, Sogang University, Seoul , Korea, Tel: , Fax: , ecsjhur@sogang.ac.kr. ** Department of Economics, Korea University, Seoul , Korea, Tel: , Fax: , cbpark_kjs@korea.ac.kr. Corresponding Author

2 I. Introduction Although it is taught in many undergraduate textbooks that openness to international trade bolsters economic growth, 1 the question as to whether a positive relationship exists between free trade and economic growth has posed a significant challenge to economists, both theoretically and empirically, since Adam Smith. On the theoretical side, the so-called endogenous growth theories hold to the proposition that trade liberalization or greater openness may promote long-run economic growth under certain conditions. For example, Grossman and Helpman (1991) and Feenstra (1996) predicted that if a free trade system is formed under conditions in which technology transfer occurs between the involved economies, production efficiency can be improved, and thus free trade can ultimately induce economic growth among the Free Trade Agreement (FTA hereafter) signatory countries. Another theoretical link between trade and growth was described in a learning-by-doing model, as emphasized by Lucas (1988) and Young (1991). If free trade allows countries to specialize in industries with economies of scale, then their long-run economic growth can be increased. These examples demonstrate that certain economic conditions are required in order to realize a positive relationship between free trade and economic growth; thus, it can be inferred that the theoretical models do not necessarily yield an unambiguous prediction regarding the relationship between free trade and economic growth. Whether free trade does indeed exert a positive effect on economic growth, whether the conditions of the models with technological spillovers are met in practice, or whether the conditions of the learning-by-doing models are supported by data are all questions that must be empirically explored. On the empirical side, however, the consensus has not been made either, and it has proven extremely difficult to discover statistical evidence of a 1 For example, Mankiw s popular textbook (2007), states on page 224: The overwhelming weight of the evidence from this body of research is that Adam Smith was right. Openness to international trade is good for economic growth. 1

3 consistent relationship between trade and economic growth. In reviewing literature from the 1990s onwards in Section II, we note that, whereas Dollar (1992), Sachs and Warner (1995), Edwards (1998), and Frankel and Romer (1999) reported supportive evidence of a positive impact of free trade on economic growth by using a variety of measures of openness, Harrison (1996), Rodriguez and Rodrik (2000), and Wacziarg and Welch (2008) found a negative effect from free trade on economic growth. In this paper, we also attempt to characterize empirically the effects of free trade on economic growth, but our approach differs from those in previous studies in two important ways. First, in the existing literature, trade openness is frequently constructed as an index reflective of the trade liberalization regimes or policies of an individual country. However, the concept of openness is hard to define and the indices are generally highly correlated with other economic variables of the country. In addition to these problems, countries with the same levels of openness may experience different effects of that openness in terms of their economic growth, depending on the openness level of their respective principal trade partners. In our paper, we focus on bilateral FTA systems among a variety of relevant trade policies to reduce mutual trade barriers and stimulate trade volume. As a bilateral FTA is formed by a pair of countries, it can be considered an indicator of mutual openness for the countries specifically involved in the agreement. The direct effects of free trade could be identified more clearly with the concept of mutual openness than with the individual trade openness referenced in earlier empirical literature regarding economic growth and trade. In analyzing these effects, we utilize a binary dummy variable to indicate whether or not a country-couple has an FTA, and consider the growth rates of both economies involved in the formation of the FTA. Although we do not deny the utility of the unilateral trade openness measures used in the existing literature, we believe that the mutual openness measure is more consistent with trade theories, where the effect of free trade is usually discussed for both economies 2

4 participating in a free trade system. As a matter of fact, the number of effective FTAs in the world has risen rapidly since the mid 1990s, as is shown in Figure 1. 2 Second, we propose a nonparametric matching approach to assess the effects of FTAs on growth. The majority of empirical studies, including the aforementioned ones, have adopted the parametric linear model using either cross-sectional or panel data. However, this linear regression approach may be subject to misspecification problems owing to potential nonlinear relations among variables, as well as the non-random selection problem. We consider engaging in an FTA as a treatment in the terminology of matching literature, and thus apply a non-parametric matching approach. This methodology imposes no parametric restrictions, and has been shown to perform well even in the face of the non-random selection problem. Although this approach is popular in labor economics, economists in the trade literature have now begun to use this econometric approach. Relevant examples include Chang and Lee (2007) and Baier and Bergstrand (2009). To the best of our current knowledge, however, this approach has never been utilized in the growth literature. 3 Our paper is organized as follows. We provide a brief literature review in Section II. In Section III, we provide a brief description of the dataset utilized in this study. After showing that FTA has a significantly positive effect on economic growth under the usual linear panel regression, we also present evidence that this linear panel regression model suffers from misspecification and non-random selection problems. Section IV briefly discusses the econometric methodology used in this study-namely, the nonparametric matching approach. In Section V, we present our main findings through nonparametric matching analysis, showing that FTAs exert no statistically significant effects on total 2 Many economists have begun to pay attention to this new phenomenon, and have conducted research on a variety of FTA-associated issues. For example, Baier and Bergstrand (2004, 2007) examined the effects of FTAs on international trade flow, using a binary FTA dummy variable in the so-called gravity models. Sohn and Lee (2006) evaluated the effects of FTA on convergence in a simple neoclassical growth model. 3 Imbens (2004) provides an excellent survey on the nonparametric matching approach. 3

5 economic growth from the one-year to ten-year horizon after launch. Furthermore, we also find an upward trend in the gap between the growth rates of per capita GDP among countries participating in an FTA. This finding implies that some countries appear to enjoy a positive FTA effect on economic growth, whereas their counterparts in the FTA may experience a negative FTA effect on economic growth. This finding may be considered to explain, in part, the insignificant effects of FTA on the combined economic growth of countries involved in an FTA. We also discuss an additional issue related to the insignificant FTA effect. Section VI presents our concluding remarks. II. Literature Review Most countries involved in FTAs anticipate the achievement of two objectives; trade promotion and an increase in economic growth. The trade effects of FTAs have been positively detected in the relevant trade literature by a number of empirical studies, notable examples of which include the works of Baier and Bergstrand (2007, 2009), Carrere (2006), Coulibaly (2007), Eicher et al. (2007), and Magee (2008). On the contrary, the desired positive effect on economic growth is generally less supported in the literature concerning economic growth. It has proven quite challenging to uncover quantitative evidence regarding a consistent relationship between trade and economic growth, and this is principally attributable to the difficulty of finding a good index of trade openness. Here, we review some empirical studies of economic growth and openness conducted since the mid-1990s, which was a period during which FTAs proliferated at a rapid pace. 4 Advocates of the theoretical hypothesis of a positive relationship between openness and growth include, among others, Dollar (1992), Sachs and Warner (1995), Edwards (1998) 4 In this review section we leave out research articles published in 1980s. The reason why we focus on the literature since mid 1990s, is because the period shows a surge of FTA formations, which is our main research subject of this paper. For a short summary of the papers in 1980s, see Harrison (1996), for example. 4

6 and Frankel and Romer (1999), all of whom utilized a variety of openness measures. Dollar (1992) constructed two indices of openness; an index of real exchange rate distortion and an index of real exchange rate variability. He argued that these two indices are reflective of the restrictiveness of trade policy, and may be negatively correlated with economic growth. Dollar then regressed the growth of real GDP per capita on these indices over the period in a sample of 95 developing countries, and reported that these indices evidenced negative and highly significant coefficients. Sachs and Warner (1995) constructed a dummy index of openness that indicates the restrictiveness of a trade regime using the following five criteria; average tariff rates, non-tariff barriers, socialist economic system, state monopoly of major exports, and black market premium. They ran a regression of growth of real GDP per capita on the dummy variable plus many other controls over the period with a sample of 111 countries and demonstrated that the openness index was, indeed, positively related with economic growth. Edwards (1998) recognized the difficulty inherent to the definition of satisfactory openness indices and attempted to test the robustness of the previous results using different existing indicators. He ran a weighted least square regression of total factor productivity on nine alternative indicators 5 of openness over the period in a sample of 93 countries. He verified a quite consistent outcome on the relationship between openness and growth. Frankel and Romer (1999) defined openness in terms of the trade-gdp ratio. The openness index was instrumentalized first by estimating a gravity equation and then by aggregating the fitted trade values. With the 1985 data from 63 countries, they regressed income per capita on the instrumentalized trade-gdp ratio and country size variables. After controlling for the endogeneity of trade with respect to income, they reported a positive causality from openness to economic growth. 5 The nine indicators are: Sachs and Warner openness index, World Development report outward orientation index, Leamer s openness index, average black market premium, average import tariff on manufacturing, average coverage of non-tariff barriers, the Heritage foundation index of distortions in international trade, collected trade taxes ratio, and Wolf s index of import distortions. 5

7 Despite all of these empirical studies that demonstrated the positive effects of trade on growth, Rodriguez and Rodrik (2000) cast some doubt on the use of the openness indices defined in the above literature. In fact, Harrison (1996) and Wacziarg and Welch (2008) have uncovered some evidence that free trade may have a negative effect on economic growth. Harrison (1996) evaluated the relationship between trade policy and growth in a countryfixed effect model. She utilized seven indices 6 of trade policy and determined that only three of the seven variables manifested the positive effect. Wacziarg and Welch (2008) extended the data of Sachs and Warner (1995) to the 1990s for 118 countries. They found that, although the average effect of the openness dummy variable on economic growth was positive, approximately half of the countries experienced zero or even negative changes in growth after trade liberalization. Whereas the aforementioned hypothetical relationship between trade liberalization and economic growth is principally concerned with whether or not an open country grows faster than a closed economy, there has been another related and intensive research conducted regarding trade liberalization and income convergence, wherein the question is whether or not an open poor country grows faster than an open rich country. However, the empirical evidence in this regard is, again, ambiguous. Ben-David (1993) determined the income convergence among the constituent countries of the European Community. More explicitly, by dividing countries into open and closed economies, Sachs and Warner (1995) and Ben- David (1996) both confirmed that the convergence is currently observed only among open countries engaging in international trade. However, Slaughter (1997) called for a careful interpretation of these earlier works, owing to the ambiguity of the concept of openness as then defined. He argued that there was a need to document carefully the exact mechanism by 6 The seven indices are calculated from several sources, such as the exchange rate, commercial policies, tariffs and nontariff barriers, black-market premium, share of trade in GDP, international prices of tradables, price distortion index, indirect bias against agriculture from industrial sector protection, and overvaluation of the exchange rate. 6

8 which trade induces convergence through factor prices, factor quantities, and production technology. In fact, Slaughter (2001) reported the opposite result-namely, that trade liberalization induced income divergence among open countries. One common fact noted throughout the literature is that there has been some difficulty inherent in the construction of a favorable index for the trade openness of an individual country. Furthermore, it appears to us that the linear regression model approach suffers from some misspecification problems. Our departure from the earlier works is that we utilized a bilateral FTA as an indicator of mutual openness which represents the trade openness of a country couple, and that we employed a nonparametric approach that imposes no specific restriction in the relation. With these strategies, we evaluate the effect of FTAs on economic growth and the issue of convergence among FTA countries. Our principal findings are, first, that trade liberalization occurring via the formation of bilateral FTAs brings about no significant effects on member countries combined economic growth, and second, that there is a divergent and uneven FTA effect on the economic growth among the member countries. III. Data and the Misspecification Test of the Linear Regression We describe the data set in this section and provide evidence of misspecification problems occurring in linear regressions. 1. Dependent Variable or Response Variable In order to estimate the effects of FTAs on the growth performance of a county couple which has an FTA, we utilize the growth rate of real gross domestic product (GDP) per person of a country couple for the dependent variable in the regression analysis, or the response variable in the matching analysis. The per capita GDP for a couple composed of Country A and 7

9 Country B is constructed as follows: where and. 2. Control Variables or Covariates As we wished to estimate the treatment effect of a bilateral FTA via the nonparametric matching approach, we required covariates that render a treated couple (countries engaged in a bilateral FTA) and an untreated couple comparable in terms of their potential growth performance and the likelihood of forming a bilateral FTA. Similarly, we also required variables to control factors that may influence the growth performance and the possibility of forming a bilateral FTA in the regression analysis. Among many variables which have been reported to have strong correlations with growth in empirical studies, we utilize 18 variables that were significantly and robustly correlated with growth in Sala-i-Martin, Dopplehofer, and Miller (2004). In addition to these 18 variables, we include 7 variables known to affect the possibility of forming a bilateral FTA, according to Baier and Bergstrand (2004) and Egger and Larch (2008). 7 The list of these variables, their brief descriptions, and the data sources are provided in Table 1. The number of countries for which these 25 variables are available is 88, and this list is provided in Table 2. In addition to these variables, year dummies for t = 1971,, 2003 are also included in the analysis. We attempt to construct a panel dataset of these variables for individual countries from 1971 to 2003 with an annual frequency. Variables that can be considered as constant over time, such as the East Asia dummy, fraction of tropical area, sub-saharan African 7 Due to the large number of countries employed in our analysis, we could not use the same definition of the variables as was used by Baier and Bergstrand (2004). Rather, we decide to borrow the similar but more relaxed definitions utilized by Egger and Larch (2008). This enables us to increase the number of countries for our analysis to 88. 8

10 dummy, Latin American dummy, Spanish colony dummy, and ethno-linguistic fractionalization are taken directly from Sala-i-Martin, Dopplehofer, and Miller (2004). 8 However, variables including the population density in coastal areas, malaria prevalence, fraction of Confucians, fraction of Muslims, and fraction of Buddhists could vary over time, but are not available in their annual frequency dataset. Assuming that there is a steady trend in these variables, we interpolate and extrapolate them using their values at two time points one is in the 1960s and the other is in the 1990s after obtaining the data for these variables from the sources as described in Table 1. All other variables are taken as annual frequency from the data sources provided in Table 1. Once the panel dataset for 88 countries was obtained, we re-construct these covariates for all 3828 (=(88ⅹ87)/2) couples over the sample period. For the majority of variables which have continuous values, the newly rebuilt covariates for a couple of countries are the weighted average of the counterparts in that couple like the above combined per capita GDP. For the binary dummy variables, the values of a dummy variable for individual countries are added for a couple. Thus, if both countries (one country) in a couple are (is) located in Latin America, then the Latin American dummy variable for this couple is indicated as two (one). If neither country in a given couple is located in Latin America, then this dummy variable is set to zero for that couple. Finally, the FTA dummy variable for a pair of countries is collected and re-arranged, on the basis of the WTO s report of Regional Trade Agreements Notified to the GATT/WTO by Date of Entry into Force for the period In Table 2, we also provide the list of Regional Trade Agreements (RTAs) that we utilize for the bilateral FTA dummy variable. The RTAs we used in the analysis are only free trade agreements and customs unions listed in the WTO. We exclude service agreements and 8 is the web address from which the data are obtained. 9

11 preferential partial agreements. We also exclude RTAs for the countries that are not in the list shown in Table 2. Among the 88 countries in the list, the number of countries belonging to at least one RTA is Misspecification Test of the Linear Regression First, we assess the fixed effect models and summarize the results in Table 3. As shown, the estimated coefficient of the FTA variable under the fixed effect model is significantly positive, which implies that FTA raises the annual growth rate of the combined per capita GDP by 0.5% per year after controlling for important factors that can affect growth and FTA formation. However, this regression has potential econometric problems, as the relation between the growth rate and the FTA membership may not be linear. The regression equation specification error test of Ramsey (1969) is employed to determine whether nonlinear transformations of the control variables have any explanatory power with regard to the dependent variable. The square of the fitted value of the dependent variable is added to the original regression, and a significant coefficient for this newly added variable can be interpreted as indicative of the potential misspecification problem of the original linear regression. It can be seen in the second column of Table 3 that the estimated coefficient of this new variable differs significantly from zero, thereby suggesting that the original linear regression has evidence for misspecification at a conventional significance level. We also re-run the model with interaction terms between the FTA variable and other control variables. We show in the third column of Table 3 that the effect of the FTA turns out to be significantly negative and many interaction terms are significant Regional Trade Agreements (RTAs); this can be considered another piece of evidence supporting misspecification. Second, although the fixed effect model has been shown to provide more consistent estimates in the presence of omitted couple-specific factors, we should exclude time-invariant 10

12 variables (e.g. East Asia dummy, sub-saharan Africa dummy, etc.) from the analysis. Considering this problem, we also investigate the potential presence of non-linearity in the random effect model as well without excluding the time-invariant variables, and the results are provided in the fourth column of Table 3. As is the case in the fixed effect model, we are able to find a significantly positive FTA effect on growth in the original random effect model. However, the Ramsey test again indicates a possible nonlinearity problem in the original regression when the square of the fitted value of the dependent variable is added. The inclusion of the interaction terms between the FTA variable and the other control variables also suggests a potential non-linearity problem in the original regression, as we were able to find several significant interaction terms. In addition to the nonlinearity problem referenced above, the linear regression may suffer from the non-random selection problem in which covariates are systematically correlated with the FTA (the treatment variable). We can think of two sources that can induce this non-random selection problem. First, the treatment variable, FTA, may exert a larger effect under certain covariate values. For example, two countries that were colonized by the Spanish in the past may see a larger effect from FTAs because the similar legal system, common language, and close cultural background between these countries can be expected to amplify the effects of FTAs on growth. Second, the treatment variable may pick up omitted non-linear relations between the dependent variable (the growth rate of combined GDP) and covariates. Regardless of the source of the non-random selection problem, the existence of this problem also induces a bias in the estimates obtained from linear regressions. Table 4 shows a list of variables for which the null hypotheses that the means of covariates across two groups (one which has an FTA and the other one which has no FTA) are equal have been rejected at the 5% significance level. As is shown in Table 4, country couples with FTA systems report significantly higher primary school enrollment rates, higher per 11

13 capita GDP, smaller tropical area, lower malaria prevalence, longer life expectancy, and lower weight of mining industry in GDP. They are also less likely to be located in sub-saharan Africa or Latin America, have higher openness measures, and also evidence lower ethnolinguistic fractionalization. In summary, the econometric problems arising from nonlinearity and non-random selection appear to warrant serious consideration. Previous studies regarding the relationship between trade and growth may not have been able to reach a consensus because the linear regression approach is vulnerable to these problems. As a consequence, we employ the nonparametric matching approach in subsequent sections in order to estimate the treatment effect of bilateral FTAs. The nonparametric matching approach can be free from the potential nonlinear problem because it imposes no specific functional forms between variables. The nonparametric matching approach is also known to perform better in the presence of the nonrandom selection problem, because it compares observations with similar characteristics. IV. Econometric Methodology Suppose that there are N countries in the dataset and those countries are indexed by i = 1, 2,, N. Y ijt denotes the combined per capita GDP for countries i and j in year t, which is the population-size weighted average of the real per capita GDP between the two countries. The growth rate of Y ijt is y ijt = ln(y ijt ) - ln(y ijt-1 ), and y ijt is the response variable in the terminology of the matching literature. A vector of covariates that contain information regarding the characteristics of a country couple (i, j) is denoted as x ijt. 9 The FTA dummy variable d ijt is one in cases in which a country couple (i, j) has a bilateral FTA at time t, and is zero in all other cases. Hence, observations with d ijt = 1 belong to the treatment group, whereas 9 In this paper, the word couple refers to an observation unit which consists of two countries in which the response variable and covariates are combined. The word pair means two observation units (couples) which are matched conditionally on the covariates. 12

14 observations with d ijt = 0 belong to the control group. Each observation in our dataset can be written as the triple (y ijt, x ijt, d ijt ) for i = 1,, N, j = i+1,, N, and t = 1,, T. Suppose that a country couple (i, j) has formed a bilateral FTA. Then, the bilateral FTA effect on the growth of the per capita GDP for couple (i, j) is defined as the difference in the growth rates when the couple forms the bilateral FTA and when the couple does not form the bilateral FTA. 10 Hence, assuming that the potential untreated response is independent of the treatment, the average FTA effect can be written as: where is the expectation operator, denotes the potential treated response, and is the potential untreated response. However, as we cannot observe 1, we need a counterfactual potential untreated response of the treated couple. The counterfactual potential untreated response of the treated couple is obtained by the matching methodology. In other words, a couple from the control group, which is the closest to the couple of the treatment group in terms of a function of the vector of covariates (x) is selected as the counterfactual potential untreated response of the treated, under the assumption that the potential untreated response occurs independently of the treatment conditional on the covariates. This assumption implies that a treated couple and an untreated couple are comparable in terms of their potential untreated growth performance and the probability of forming bilateral FTA if they evidence the same covariates. When we select a couple,, from the control group as the counterfactual for a given couple,, from the treatment group, we employ two functions of covariates as criteria. The first function of covariates is the Mahalanobis metric, 10 In terms of the terminology of the matching literature, we are interested in estimating the effect on the treated. The method in this paper can also be applied to estimate the effect of the untreated. 13

15 which is defined as where is the sample variance matrix of covariates in the pooled sample. The Mahalanobis metric is a scale-normalized distance between and, and the couple,, from the control group which minimizes this scale-normalized distance is selected as the counterfactual potential untreated response of the treated couple (i, j). The second function is the propensity score function, which can be interpreted as the estimated probability of forming an FTA as a function of a country couple s covariates. The probit model is utilized to estimate the propensity score. For each treated couple (i, j), only one untreated couple which has the closest estimated propensity score is selected as the counterfactual. While searching for the most appropriate untreated couple for each treated couple, we use some control couples more than once in matching. 11 Once each treated couple is matched with its corresponding untreated couple by either of these functions, the average effect on the treated is nonparametrically estimated. That is, without assuming any particular functional form, the estimated average effect of the FTA is computed as the average difference of the response variable (the growth rate of combined GDP per capita) across matched pairs. Although we select an untreated couple that minimizes the scale-normalized distance or has the nearest propensity score to a given treated couple as a matching partner, the minimized distance or the nearest propensity score could be rather large in certain cases, which means that there could be no good matching for a certain treated couple. In an effort to avoid such cases, we can discard the matched pairs falling within the worst % of all matched pairs We use the psmatch2 command developed by Leuven and Sianesi (2003). 12 The cost of discarding the matched pairs whose distance falls within the worst % of all matched pairs is the reduction in the number of observations. However, the results reported in subsequent sections are robust when we utilize all matched pairs and when we further restrict the sample by discarding the worst 20% of all matched pairs in terms of the distance or propensity score. 14

16 V. Empirical Results 1. FTA Effect on Economic Growth This section presents empirical results regarding the effects of FTA on the growth of the per capita GDP of treated couples using the matching method. For each treated couple (a country couple that has an FTA), one country couple from the control group (country couples which have no FTA) is matched. The matching criterion involves selecting the one with the minimum Mahalanobis metric or the nearest propensity score among couples in the control group. Using these matched pairs, we estimate the effects of FTA on the growth of treated couples. Table 5 shows the results concerning the effects of FTA on the annual growth rates of treated country couples. Results based on both the Mahalanobis metric and propensity scores are provided. As is shown in Table 5, we estimate the effect with two caliper choices: 100% and 80%. The caliper choice of 100% means that we have utilized all matched pairs to estimate the effect, whereas the caliper choice of 80% means that we estimate the effect after discarding the worst 20% of matched pairs in terms of the Mahalanobis metric or the propensity score. The total number of matched pairs for the 100% caliper choice is The results in Table 5 are remarkably different from those obtained from the linear panel regression in Section III. The results based on the Mahalanobis metric states that the estimated effect of FTA on the growth of treated couples is insignificantly negative regardless of the caliper choices. The results based on the propensity score are similar. As the results in Table 5 show the average difference in the annual growth rate of per capita GDP between country couples with and without an FTA, the observations include the FTA effect in the very short-run (e.g. the difference in the annual growth rate one year after the FTA has started) as well as the FTA effect in the very long-run (e.g. the difference in the annual growth rate more than twenty years after the FTA has started). However, the short-run 15

17 FTA effect could differ substantially from the long-run one. For example, it may take a long time for some industries to realize the benefit of an FTA, whereas it may take a relatively short time for other industries to be negatively affected by the FTA. Hence, the effect of the FTA in the short-run tends to be negative or insignificant. However, as workers and resources move gradually to industries that have been affected positively by the FTA, the FTA effect may, in the long-run, become more positive. As the number of country couples engaged in FTAs has increased rapidly since the mid-1990s, the results in Table 5 might be biased toward the short-run effect. In an effort to overcome this problem and to understand the dynamics of the FTA effect over time, we assess the cumulative growth rates of matched pairs since the time at which the treated couple in a given pair formed their FTA. The number of matched pairs is greatly reduced and varies from 326 (the number of matched pairs one year after the launch of FTA) to 122 (the number of matched pairs ten years after the launch of FTA) when the caliper choice is 100%. Figure 2 shows the average difference in the cumulative growth rates of matched pairs since the time at which the treated couple in a given pair forms an FTA. Although the caliper choice in Figure 2 is 80%, the results are quite similar to the other caliper value choice (100%). 13 Figure 2 shows the 95% confidence interval for the FTA effect. As is shown in the left panel of Figure 2, the FTA effect on the cumulative growth rates is positive in most years from one year to ten years after the beginning of the FTA on the basis of the Mahalanobis metric, but is always insignificant. Furthermore, no upward sloping trend at all was noted with the progression of the time horizon. The results based on the propensity score are generally consistent with those based on the Mahalanobis metric, except that the estimated effect fluctuates more profoundly. In summary, the results in Figure 2 indicate that the formation of an FTA does not significantly stimulate the growth performance of countries 13 The results with the caliper value of 100% are available upon request. 16

18 involved, up to a ten-year horizon. 2. FTA Effect on Growth Rate Convergence Although many countries have recently formed FTAs or have begun to consider forming FTAs, the results shown in Table 5 and Figure 2 imply that the average effect of FTA on the growth rate of per capita GDP for country couples with FTAs is insignificant. In this subsection, we attempt to determine whether this insignificant effect is the consequence of the convergence of per capita GDP growth rates of individual countries in a treated couple, or of the divergence of per capita GDP growth rates of individual countries in a treated couple. In other words, we attempt to determine whether the gap in the per capita GDP growth rates of two countries with an FTA has grown or declined since the inception of the FTA. The nonparametric matching approach is utilized to compare the differences in the growth rates between country couples with and without an FTA, after switching the response variable from the growth rate of combined per capita GDP to the difference of growth rates across two countries within a couple. The results are provided in Figure 3. As is shown in Figure 3, the gap of the growth rates between members of a country couple with an FTA is initially lower on average than that between members of a country couple without an FTA immediately following the launch of the FTA by both matching methods. That is, an FTA seeks to lower the gap of the growth rates between member countries in a couple with the FTA, compared with that between member countries in a couple without the FTA, although this effect is insignificant. As time elapses, however, the gap of the growth rates between member countries in a couple rises relatively more rapidly when these two countries have formed an FTA. The gap between countries with an FTA becomes significantly higher than that between countries without an FTA five and six years (ten years) after the start of FTA, based on the Mahalanobis metric (the propensity score). The upward trend and significantly 17

19 positive gap shown in Figure 3 suggest that the effects of FTAs on economic growth may not be symmetrical between country members within an FTA. Some countries appear to enjoy a positive FTA effect on economic growth, whereas their FTA partners appear to experience a negative FTA effect on economic growth. 3. The Behavior of Growth Rates Before FTAs One may surmise that we may have found the insignificant FTA effect on the combined growth rate of GDP because the GDP growth rates for couples from the treatment group were lower than the matched couples from the control group before the inception of the FTAs. In other words, when we match a couple from the treatment group with a couple from the control group on the basis of the Mahalanobis metric or the propensity score, we implicitly assume that the growth behaviors between the matched couples are similar prior to the inception of the FTAs; this assumption may, however, not be true. In order to evaluate this possibility, we attempt to determine whether any significant difference exists in the behavior of combined GDP growth rates between matched couples during a period prior to the formation of FTAs. As it generally requires approximately one or two years for countries to complete their FTA negotiations and to put their FTA into effect, we compare three- to five-year cumulative growth rates prior to the launch of FTAs for matched couples in order to determine whether those growth rates were comparable prior to the formation of FTAs. The T-statistics of the difference in the cumulative growth rates of combined GDP are presented in Table 6. As is shown in Table 6, the differences in the cumulative growth rates of combined GDP between matched couples are insignificant in all cases, except that the difference in the five-year cumulative growth rates is marginally significant at the 5% level when the propensity score is utilized as the matching criterion. These results indicate that the 18

20 insignificant effect of the FTA on the combined growth rates is not attributable to a systematic difference in the combined growth rates between matched couples prior to the inception of FTAs. VI. Concluding Remarks In this paper, we utilize an alternative trade openness measure which implies mutual trade liberalization and the nonparametric matching approach to determine whether an FTA exerts an effect on the growth of two countries involved in the FTA. Whereas linear panel regressions, which are common in both the growth literature and the FTA literature, are found to be subject to econometric problems including misspecification and non-random selection, the nonparametric matching method imposes no specific functional form in the relation, and can be applied to a broad range of data structures. The principal finding of this paper is that FTAs were determined to exert an insignificant effect on the growth performance from the nonparametric matching approach. An FTA has an insignificant effect in the one-year to ten-year horizon after the launch of the FTA. Furthermore, we find an upward trend in the gap between the growth rates of per capita GDP among countries within an FTA. This is suggestive of uneven FTA effects across countries within an FTA, which may explain, in part, the insignificant effects of FTAs on the economic growth of both countries within an FTA. This also indicates that policy makers should not consider FTAs as a strategy that guarantees rapid development. One may surmise that the insignificant effects of an FTA on the combined growth rate of GDP might be attributable to certain distortions induced by FTAs. FTAs are basically preferential free trade agreements among member countries, and this preferential arrangement can exert a trade diversion effect against non-fta member countries outside the FTA blocs. If the trade diversion effect is relatively more dominant, then the effect of FTAs 19

21 on the growth rate might prove insignificant, and this may explain why we detected an insignificant FTA effect. However, recent studies such as that of Lee et al. (2008) have already demonstrated that not only insiders of RTAs but also outsiders increase their trade among countries. According to this empirical finding, the trade diversion effect of FTAs would not prove a compelling reason for the insignificant effect of FTAs on economic growth. Although our findings appear remarkable in relation to the conventional view regarding the effects of free trade, our results can be considered consistent with recent endogenous growth models that focus on the relationship between trade and growth. For example, Feenstra (1996) has shown that the free trade system does not increase the growth rate of large countries, and even slows down the growth rate of small countries in the long run. This theoretical proposition implies an insignificant FTA effect on combined growth rates and an uneven FTA effect between the countries involved in an FTA, which are consistent with our empirical findings in this paper. 20

22 References Baier, Scott L. and Jeffrey H. Bergstrand, 2004, Economic Determinants of Free Trade Agreements, Journal of International Economics 64(1), pp Baier, Scott L. and Jeffrey H. Bergstrand, 2007, Do Free Trade Agreements Actually Increase Members International Trade?, Journal of International Economics 71(1), pp Baier, Scott L. and Jeffrey H. Bergstrand, 2009, Estimating the Effects of Free Trade Agreements on International Trade Flows Using Matching Econometrics, Journal of International Economics, 77(1), Ben-David, Dan, 1993, Equalizing Exchange: Trade Liberalization and Income Covergence, Quarterly Journal of Economics, 108(3), pp Ben-David, Dan, 1996, Trade and Convergence among Countries, Journal of International Economics, 40(3-4), pp Chang, Pao-Li and Myoung-Jae Lee, 2007, The WTO Trade Effect, Manuscript. Carrere, Celin, 2006, Revisiting the Effects of Regional Trade Agreements on Trade Flows with Proper Specification of the Gravity Model, European Economic Review, 50(2), pp Coulibaly, Souleymane, 2007, Evaluating the Trade Effect of Devloping Regional Trade Agreements: A Semi-Parametric Approach, World Bank Policy Research Working Paper Dollar, David, 1992, Outward-Oriented Developing Economies Really Do Grow More Rapidly: Evidence from 95 LDCs, , Economic Development and Cultural Change 40(3), pp Edwards, Sebastian, 1998, Openness, Productivity and Growth: What Do We Really Know? Economic Journal 108(447), pp Egger, Peter and Mario Larch, 2008, Interdependent Preferential Trade Agreement Memberships: An Empirical Analysis, Journal of International Economics, 76(2), pp Eicher, Theo, Christian Henn, and Chris Papageorgiou, 2008, Trade Creating and Diversion Revisited: Accounting for Model Uncertainty and Natural Trading Partner Effects, IMF Working Paper WP/08/66. Feenstra, Robert C., 1996, Trade and Uneven Growth, Journal of Development Economics 49(1), pp Frankel, Jeffrey A., and David Romer, 1999, Does Trade Cause Growth? American Economic Review 89(3), pp

23 Grossman, Gene M. and Elhanan Helpman, 1991, Innovation and Growth in the Global Economy, MIT Press, Cambridge, MA. Harrison, Ann E., 1996, Openness and Growth: A Time-Series, Cross-Sectional Analysis for Developing Countries, Journal of Development Economics 48(2), pp Imbens, Guido W., 2004, Nonparametric Estimation of Average Treatment Effects under Exogeneity: A Review, Review of Economics and Statistics 86(1), pp Lee, Jong-Wha, Innwon Park, and Kwanho Shin, Proliferating regional trade arrangements: Why and Whither?, The World Economy, 31(12), pp Leuven, Edwin and Barbara Sianesi, 2003, PSMATCH2: Stata module to perform full Mahalanobis and propensity score matching, common support graphing, and covariate imbalancing testing, Statistical Software Components, S432001, Department of Economics, Boston College. Lucas, Robert E., Jr., 1988, On the Mechanics of Economic Development, Journal of Monetary Economics, 22(1), pp Magee, Christopher S. P., 2008, New Measures of Trade Creation and Trade Diversion, Journal of International Economics, 75(2), pp Mankiw, N. Gregory, 2007, Macroeconomics, 6 th edition, Worth publishers. Ramsey, J. B. (1969) Tests for Specification Errors in Classical Linear Least-Squares Regression Analysis, Journal of the Royal Statistical Society (Series B), 31:2, Rodriguez, Francisco, and Dani Rodrik, 2000, Trade Policy and Economic Growth: A Skeptic s Guide to the Cross-national Evidence, in Ben S. Gernanke and Kenneth Rogoff, eds., NBER Macroeconomics Annual, 2000, MIT Press, Cambridge, Massachusetts. Sachs, Jeffrey D., and Andrew Warner, 1995, Economic Reform and the Process of Global Integration, Brooking papers on Economic Activity 1, pp Sala-i-Martin, Xavier, Gernot Doppelhofer and Ronald I. Miller, 2004, Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach, American Economic Review 94(4), pp Slaughter, Matthew J., 1997, Per Capita Income Convergence and the Role of International Trade, American Economic Review, 87(2), pp Slaughter, Matthew J., 2001, Trade Liberalization and Per Capita Income Convergence: A Difference-in-Difference Analysis, Journal of International Economics, 55(1), pp Sohn, Chan-Hyun and Hongshik Lee, 2006, How FTAs Affect Income Levels of Member Countries, The World Economy 29, pp

24 Wacziarg, Romain and Karen Horn Welch, Trade Liberalization and Growth, World Bank Economic Review, 22(2), pp Young, Alwyn, 1991, Learning by Doing and the Dynamic Effects of International Trade, Quarterly Journal of Economics, 106(2), pp

25 Table 1: Data Descriptions and Sources Variable Description Source East Asian dummy Dummy for East Asian Countries Sala-i-Martin, Dopplehofer and Miller (2004) Primary schooling Annual enrollment rate in primary World Development Indicator school Investment price Annual investment price Penn World Table 6.2 Real GDP pc Real GDP per capita Penn World Table 6.2 Fraction of tropical area Proportion of country s land area within tropics Sala-i-Martin, Dopplehofer and Miller (2004) Population density in coastal area Coastal population per coastal area Gallup et al. (2001) (interpolated) Malaria prevalence Index of Malaria prevalence Gallup et al. (2001) (interpolated) Life expectancy Annual life expectancy World Development Indicator Confucian Fraction of Confucian Population Barro (interpolated) African dummy Dummy for sub-saharan African countries Sala-i-Martin, Dopplehofer and Miller (2004) Latin American dummy Dummy for Latin American countries Sala-i-Martin, Dopplehofer and Miller (2004) Fraction of GDP in mining Fraction of GDP in mining Sala-i-Martin, Dopplehofer and Miller (2004) Spanish colony Dummy for former Spanish colonies Sala-i-Martin, Dopplehofer and Miller (2004) Openness 14 Openness variable Penn World Table 6.2 Fraction of Muslim Fraction of Muslim population (interpolated) Fraction of Buddhist Fraction of Buddhist population Barro (interpolated) Ethonoliguistic fractionalization Average of five difference indices of ethnolinguistic fractionalization Sala-i-Martin, Dopplehofer and Miller (2004) Government consumption share Annual share of government Penn World Table 6.2 consumption to GDP Natural Inverse of Distance CIA World Fact Book Remoteness Remoteness of coupled countries from CIA World Fact Book the rest of the world GDP sum Penn World Table 6.2 GDP sim log Penn World Table 6.2 DKL / Penn World Table 6.2 / SQDKL DKL Penn World Table 6.2 DROWKL 0.5{ / + / } where Penn World Table 6.2 The web address for the data source indicated as Sala-i-Martin, Dopplehofer and Miller (2004) is The web address for the data source indicated as Gallup et al. (2001) is The data for religion is taken from Robert Barro s web site of which the address is Penn World Table 6.2 can be found at World Development Indicator can be found at the web site of the World Bank which is Barro 14 The variable used in Sala-i-Martin, Dopplehofer and Miller (2004) is the number of years for which the economy has been open between 1950 and Since we need a variable with annual frequency for individual countries, we use openness variable in Penn World Table. 24