Estimating Regional Trade Agreement Effects on FDI in an Interdependent World

Recent research on trade and multinationals highlights a novel issue with multinational firms. In particular, their integration strategies are complex and the degree of vertical integration varies in a multilateral world with many possible locations of activity. Multinationals may choose some plants to serve consumers locally only, whereas others engage in trade. Overall, this may explain the fact that a high percentage of world trade is actually controlled by multinational firms, although most of the foreign direct investment (FDI) occurs within the block of developed countries. The most important regional trade agreements (RTAs) are signed between members of the very same block of economies. This gives rise to the question asked in the present paper: what is the impact of RTAs on FDI in an interdependent world? The paper focuses on the role of the Europe Agreements between the member countries of the European Union and ten Central and Eastern European countries. In doing so, recent spatial HAC estimation techniques are applied to both estimation and testing.


Introduction
The second half of the last century was characterized by a surge of "bilateralism" in trade policy. The foundation of the European Union (EU, formerly referred to as European Community), the European Free Trade Area (EFTA), and the North American Free Trade Area (NAFTA) are some of the most sizeable regional trade agreements (RTAs) that were signed and implemented within this period. As observed by authorities in empirical research on trade issues, this process resulted in a significant increase in bilateral trade volumes among the member countries (see Baier andBergstrand, 2007, or Glick andRose, 2002). At the same time, foreign direct investment (FDI) increased much faster than trade, even within the OECD and among the members of the mentioned RTAs. While numerous studies on the impact of RTAs on bilateral trade are now available, the question of how the bilateralism of trade policy affects FDI seems under-researched.
The theory of horizontal multinational firms (Markusen, 1984) assumes that the avoidance of trade impediments (including tariffs and other trade costs) is a major reason for setting up foreign plants that produce the same good in the parent and the host country. By way of contrast, vertical multinational firms (Helpman, 1984) split up the production process across borders to exploit gains from comparative advantage within the firm. For instance, the gains from "outsourcing" production stages to low-wage countries and the associated trade of intermediate goods within firms are important issues with vertical multinational firms. Since these firms engage in trade, we expect vertical FDI to increase through the implementation of RTAs. Hence, the sign and magnitude of the coefficient of the RTA variable (typically a dummy variable) in empirical FDI specifications is of interest for policy analysis. It also implicitly indicates the relative importance of horizontal versus vertical FDI.
However, more recent theory points to the complex integration strategies of multinational firms (Yeaple, 2003, Helpman, Melitz, and Yeaple, 2004, Raff, 2004, and Grossman, Helpman, and Szeidl, 2007. In particular, this literature avoids the restrictive features of models with simple horizontal or vertical multinationals. While it may be optimal to set up foreign subsidiaries in some host countries to serve only the local consumers (the horizontal motive), it may be optimal for the same firm to set up export platforms in other host countries that serve consumers there and elsewhere. Hence, this theory comes closer to the empirical stylized facts of mixed horizontal-3 vertical and complex integration strategies within multinationals. Two issues with complex multinational firms are of particular interest in the present paper. First, it is an empirical reality that foreign subsidiaries are set up in a multi-country world, and it is potentially insufficient to model bilateral FDI as a function of bilateral determinants only. Firms set up their foreign plants in accordance with the characteristics not only of a particular target market but also with the characteristics of other potential host countries.
Second, the design of a multinational production and sales network likely entails strategic aspects of plant location in space. 1 Third, the role of RTAs will be nontrivial with complex FDI. Low trade barriers are an incentive to export not only for national but also for complex multinational firms (similar to vertical multinationals). However, high trade barriers foster the location of locally selling foreign subsidiaries (similar to horizontal multinationals).
Overall, the net effect of a reduction in trade barriers is less clear-cut in complex than in simple forms of the multinational firm organization.
How does empirical work on the impact of RTAs relate to the theory of multinational firms? As mentioned above, only a few articles address this issue. Blomström and Kokko (1997) report on three case studies. They point out that the implementation of the U.S.-Canada Free Trade Agreement led to a reduction in intra-regional FDI to both the U.S. and Canada (i.e., a negative impact on bilateral FDI), while it increased extra-regional FDI into  (2000), Blonigen, Davies, Waddell, Naughton (2005, 2007, and Baltagi, Egger, and Pfaffermayr (2007), using different data, and spatial cross-sectional as well as spatial panel data models, find that FDI between two countries is not independent of FDI in other economies. This is expected from a general equilibrium perspective (see Blonigen, 2005, for a survey). In a similar vein, Drukker and Millimet (2007) Blonigen, Davies, Waddell, and Naughton, 2007). Also the Europe agreement dummy is spatially weighted to capture the third country effects of trade liberalization on FDI. (ii) The disturbances are allowed to be spatially correlated because of the regional interdependencies of stochastic shocks between the host countries. Accordingly, we calculate spatial HAC robust standard errors as proposed by Kelejian and Prucha (2007).
The estimation results illustrate that third-country effects are important and lend support to a complex impact of the Europe Agreements on FDI.
The findings indicate that RTA membership of a European host country leads to a relocation of FDI from other countries to RTA members. This is consistent with models of export-platform FDI, where multinational firms (re-)locate their subsidiaries in countries from which consumers in a larger area can be served at lower trade costs (Raff, 2004;and Ekholm, Forslid, Markusen, 2007).
The remainder of the paper is organized as follows. The next section outlines the specification of bilateral FDI as supported by recent general equilibrium theory. Section 3 provides details on the adopted econometric approach. Section 4 reports the findings regarding the impact of the Europe Agreements on bilateral FDI. Section 5 provides our policy recommendations, and the last section concludes with a summary of the most important findings.

6
2 Determinants of bilateral FDI and the role of regional trade agreements According to previous research, the most important empirical determinants of multinational firm location are country size, skilled labor endowments, trade and investment costs, and interaction terms thereof (Carr, Markusen, and Maskus, 2001;Markusen and Maskus, 2002;and Blonigen, Davies, and Head, 2003). While the actually estimated models are often in levels rather than in logs, the latter approach is typically preferable from an econometric point of view, as pointed out by Mutti and Grubert (2004). Taking this into account, the log of FDI from country i to country j, y ij , may be formulated as a log-linear function of the following explanatory variables (see Markusen, 2002): 2 the sum of home and host country GDP, , the relative size of the home and the host market in terms of GDP, log GDP i − log GDP j , 3 and the difference in exporter and importer skill endowment ratios (approximated by gross secondary school enrolment in percent), DSK ij = SK i − SK j . The latter is also used in five interaction terms to account for the non-linear impact of skilled labor endowment differences on FDI: In the application, we consider a panel data set with fixed country-pair and time dummies. However, we skip the time index in the variable definition. 3 We know from previous theoretical and empirical work that bilateral FDI stocks increase in parent-to-host country relative GDP (or the corresponding log-difference; see Bergstrand and Egger, 2007). Accordingly, we employ such a specification instead of the squared difference in GDPs as used by others.
that takes on the value 1 if the condition in parentheses holds and 0 other- where DIST ij is the great circle distance between two countries' capitals, serving as a proxy for trade costs.
Whereas horizontal FDI should rise if two markets grow larger and become more similar (i.e., if SGDP ij and RGDP ij increase), vertical FDI should rise if the parent country is small and well endowed with skilled labor, and trade costs between the two markets are low. Accordingly, we expect a positive sign for the coefficients of SGDP ij and RGDP ij but a negative one for the coefficients of all skilled labor endowment interaction terms. While there is no clear-cut hypothesis for the sign of the main effect of DSK ij , we expect INT 1 ij and INT 4 ij to enter positively and the other interaction terms to exert a negative impact on FDI (see Carr, Markusen, and Maskus, 2001;and Markusen and Maskus, 2002).
Note that the sample of 23 parent and 28 host countries covers only member countries of the European Economic Area (EEA) and the ten Central and East European countries (CEEC) that have successfully applied for EU membership (see the Appendix for a detailed list of economies). Since there was no change to the composition of the EEA within the sample period, its effect is captured by the country-pair dummies. However, ten Europe Agreements between the EU and a CEEC have been ratified within the period considered.
These are the ones with Hungary and Poland in 1994;with Bulgaria, Czech Republic, Romania, and Slovak Republic in 1995;with Estonia, Latvia, and Lithuania in 1998;and with Slovenia in 1999 (see the Appendix for further 8 details). We capture these agreements by the dummy variable EA ij which takes the value 1 for two economies that ratified the Europe Agreement in a given year and afterwards. For all other country pairs and years this dummy variable is zero. Hence, this dummy variable exhibits time and country-pair variation. By controlling for country-pair and time effects, the corresponding parameter can be interpreted as a difference-in-difference direct effect of the Europe Agreements on bilateral FDI.
To simplify the exposition of our econometric approach, we collect all mentioned variables in the matrix X n = [EA n , SGDP n , RGDP n , DSK n , INT1 n , ..., INT5 n ], where X n is an n × k matrix with n being the number of observations and k = 9 denoting the number of variables collected in X n .
The Appendix gives details on the variable sources for both the dependent (log bilateral outbound FDI) and independent variables. Moreover, Tables A.1-A.3 in the Appendix provide descriptive statistics of the dependent and independent variables as well as partial correlation coefficients. We allow for spatial interdependence in X n across host countries (since FDI location decisions depend not only on the parent and actual host country characteristics but also on the characteristics of the competing European host markets).
The spatial weights matrix is defined in the following section, which serves to aggregate the characteristics of a host market's competitors. We do so to account for the possibility that the Europe Agreements not only affect FDI decisions directly in a particular new entrant but may indirectly affect FDI decisions in the other competing host markets. 9

Econometric approach
We consider the following econometric model: y n = X n α n +X n β n + D n µ n + u n (1) = Z n δ n + u n where Z n = [X n ,X n , D n ] and δ n = [α 0 n , β 0 n , µ 0 n ] 0 . To simplify notation, let us refer to a specific country-pair ij in year t by p. The total number of observations is n = P P p=1 T p , where P denotes the number of unique countrypairs and T p is number of observations when country-pair p is observed. P = MN, where M (N) denotes the number of unique parent (host) countries. y n is an n×1 vector of observations of the dependent variable (with elements y ijt = log F DI ijt ), X n is the n × k matrix of explanatory variables, including the Europe Agreements dummy variable EA n , and D n is an n × l matrix of (country-pair and time) dummy variables, where l = P + max[T p ].
We refer toX n as the spatial lag of X n (see below for further details).
While the panel data-set is unbalanced due to missing elements in y n ,X n is computed from the balanced data. IfX n were based on spatially weighted averages in the unbalanced panel data-set, we would obtain biased and inconsistent estimates of the parameters and the disturbances, even with randomly missing elements in y n . The reason for the latter is that some missing observations with a non-zero spatial weight would not be accounted for in the spatial averages, leading to measurement error. To circumvent this problem, we define X P T andX P T = W P T X P T as the balanced counterparts to X n andX n , respectively, where P T > n and W P T is a block-diagonal P T × P T spatial weighting matrix with MT blocks W N of size N × N. X n andX n are derived by eliminating the rows corresponding to missing values in the dependent variable from X P T andX P T , respectively. δ n is a (2k + l + 1) × 1 vector of unknown parameters.
Each block W N of W P T = I MT ⊗ W N refers to a specific parent country and year, and it includes the spatial weight among FDI-hosts. This weighting scheme implies that there is only spatial interdependence among the hosts of a specific parent country in a given year, while there is no such interdependence across parent countries or across time periods. The diagonal elements of W N are 0. We define the off-diagonal elements of W N as w jk /w * , for j, k = 1,..., N, i.e., j and k run over host countries. In our application, w jk corresponds to the log 'natural' trade flow (exports of j to k plus exports of k to j) in nominal U.S. dollars averaged over the period 1990 to 2000.
Log natural trade flows are defined as the predictions from a model of bilateral exports. These predictions are referred to as 'natural' trade because they reflect the systematic part of trade flows as suggested by economic theory. Interdependence is established via natural trade flows. The reasoning is that the multinational's outside option of serving consumers in that host country locally, via FDI there, is to supply goods to consumers in that country via exports from a third country (see Raff, 2004). Natural trade flows are used in W N because they reflect the trade potential between two markets as predicted by a gravity model of bilateral trade, accounting not only for geographical distance and proximity, but also for market size and other country-specific determinants of trade. 4 This weighting scheme supports the theoretical view that a given parent country's FDI in a given host country depends on this country's and the other host countries' characteristics.
In the case of row-normalization of W N , w * = P N k=1 w jk is the sum of the elements in the corresponding row of W. Under maximum row-sum corresponds to the maximum of these row-sums. In the former case, the spatial weighting matrix is row-normalized, and in the latter case it is normalized by the maximal row sum as suggested by Kelejian and Prucha (2005). u n = y n − Z n δ n is a vector of disturbances. In the estimation, we guard against possible heteroskedasticity and correlation of the disturbances across space and, alternatively, across time. While the latter can be accomplished by applying a standard HAC estimator as proposed by Newey and West (1987), spatial HAC (SHAC) estimators have been proposed by Conley (1999) and, more recently, by Kelejian and Prucha (2007). We apply the SHAC procedure of Kelejian and Prucha, since it is robust to measurement error of the spatial distance metric (in our case, natural trade flows). Furthermore, the estimator and importer-specific fixed effects and the following set of dyad-specific explanatory variables (we give the coefficients and standard errors in parentheses) in the model to predict Database, distance is measured as the great circle distance between countries' capitals, and the dummy variables are gathered from the CIA's World Factbook.
of the variance-covariance matrix is based on a set of assumptions that is satisfied in a wide class of spatial models.
To compute the SHAC variance-covariance matrix of the estimated parameters, denote the rth and sth variables in the matrix Z for a specific country-pair and time t as z or,n and z os,n , respectively. Similarly, the values with entries d * oo 0 ,n = (w oo 0 ,n /w * ) −1 . 5 Similar to X n , we assume that spatial interdependence of the disturbances only occurs across host countries within a year for a given parent country. All elements d * oo 0 ,n with different parent countries or years and all elements d * oo 0 ,n where o = o 0 are set to d * oo 0 ,n = 0. Additionally, let d n be a critical value determining the radius of spatial interdependence. Hence, spatial interdependence is only assumed for those The SHAC estimator of Kelejian and Prucha (2007) involves the kernel function K(d * oo 0 ,n /d n ). For the latter, we assume a Bartlett window which is given by Kelejian and Prucha obtain a consistent estimate for the (r, s)th element of 5 Kelejian and Prucha (2007) assume that d * oo 0 ,n = d oo 0 ,n +υ oo 0 ,n , where d oo 0 ,n is the true measure of economic 'distance' (in our case the inverse natural trade flow) between observations o and o 0 and υ oo 0 ,n = υ o 0 ot,n is a measurement error where (υ oo 0 ,n ) is independent of (ε o,n ). For convenience, we define w * in the same way as in W P T . the variance-covariance matrix which is given bŷ The SHAC-based variance-covariance matrixΨ n = (ψ rs,n ) may then be used for hypotheses testing. 6 For instance, this is useful to test the hypothesis of the joint relevance of the spatial lags in the exogenous explanatory variables in the subsequent empirical analysis.
4 Empirical analysis -the impact of the Europe Agreements on bilateral FDI in Europe For robustness, we employ two different spatial weighting schemes described above. Both of them are based on natural trade flows among host countries, but they differ with respect to the normalization method. Most of the existing applications of spatial econometric models rely on row-normalized matrices W P T . However, Kelejian and Prucha (2005) point out that it is sufficient to normalize all entries of W P T and, hence, of W P T , by the largest eigenvalue or, alternatively, by the largest row-sum of W P T . Row-normalization imposes strong restrictions on the spatial process, since each row of W P T is normalized differently (hence, only relative economic distance matters but there is no role to play for absolute economic distance). By way of contrast, the maximum row-sum normalization suggested by Kelejian and Prucha (2005) implies dividing the whole matrix W P T by a single scalar which preserves the importance of absolute economic distance along with that of relative dis- intra-firm trade. The reason is that a larger absolute economic distance between host markets renders intra-firm trade in goods more costly. However, a row-normalized weighting scheme does not support any role for economic distance in absolute terms, since it only relies on relative economic distance to other host countries in the spatial weighting scheme. To see this, suppose that one parent country's host markets exhibit an economic distance which is ten times that of another parent country. If all economic distances are the same among the host markets for a parent country, the row-normalized weighting scheme exhibits identical entries, even though absolute distances differ by a factor of ten across parent countries. This is not the case for the alternative weighting scheme suggested by Kelejian and Prucha (2005), where the absolute role of economic distance is maintained, being in line with economic theory.
> The results indicate that parent and host country joint size and parentto-host-market relative size are positively related to bilateral outward FDI   the Europe Agreements depends also on a host country's economic distance from other host markets. Hence, the overall impact of Europe Agreement membership on inward FDI (from the average parent country in the sample) will differ across host countries.
In a multi-country world, the other explanatory variables also exhibit bilateral and third-country effects. For instance, not only do bilateral relative and absolute factor endowments (i.e., country size) matter, but endowments of all competing host markets are relevant (see also Head, Ries, andSwenson, 1995, andBlonigen, Davies, Waddell, andNaughton, 2007, for the inclusion of the impact of exogenous explanatory variables of adjacent/non-distant locations of FDI). 8 In our application, all elements of X n except the spatially weighted effect of INT 5 enter significantly. This is reflected in a significant Wald statistic testing the joint significance of all elements of X n .
Overall, the opposite signs of the bilateral and the spatially weighted effects of both EA and SGDP support the hypotheses generated by models of export-platform FDI, where FDI decisions are substitutive across host markets (see Raff, 2004;and Ekholm, Forslid, and Markusen, 2007). 9 Even 8 For instance, a host market's growth is not sufficient to stimulate bilateral inward FDI. What matters is whether it grows faster or slower than the rest of the competing host countries. 9 In the sensitivity analysis summarized in the Appendix, we investigate the dynamics of the adjustment of bilateral FDI stocks to changes in the explanatory variables. We accomplish this by adding once time-lagged explanatory variables to the model. Note that it is not possible to add further lags without losing countries such as the Czech Republic or the Slovak Republic from the sample. However, the contemporaneous and the once-lagged determinants are highly correlated (see Table A.2 in the Appendix), rendering identification of the dynamic pattern difficult. For instance, the partial correlation coefficient between though some of the coefficients of the spatially weighted variables are larger in absolute value than the corresponding ones for the unweighted variables, the impact of changes in a single third host country on bilateral FDI is rather small. The reason is that we need to multiply these estimates by the third country's spatial weight. The latter is much smaller than unity for all host countries due to row-normalization of W N . However, what happens in all third host countries together may be more important for bilateral FDI than what happens in the target country. Of course, traditional explanatory variables do not necessarily take this into account, because they do not capture directly third-country effects.
In  Table A.3 in the Appendix for a summary of the results).
Of course, inward FDI of the WEC from WEC parents does not entertain any direct effect of a Europe Agreement ratification. Yet, these countries face a negative indirect impact from some of the CEEC ratifications. The largest indirect effects in Table 2 indicate the most important trading partners of the ratifying CEEC in the corresponding period. By and large, Table 2 indicates that the negative indirect impact on the average WEC host country's inward FDI from the WEC was largest in 1995 (when four agreements had been ratified with CEEC) and smallest in 1999 (when only one agreement had been ratified with Slovenia). In the same years, the CEEC had faced the largest and smallest average positive direct impact.
Note that the impact of EA ratification on an involved CEEC is much larger than that on a WEC. The reason is that the number of involved CEEC is small as compared to the number of WEC. Therefore, the overall impact on FDI into a CEEC is much larger than the one on FDI into a WEC. Also, there is no direct effect on intra-WEC bilateral outbound FDI in Table 2.
In fact, the Europe Agreements did not involve any change in trade barriers

Discussion and policy recommendations
The limited availability of resources -such as factor endowments -naturally establishes interdependence in the allocation decisions about these resources.
We provide empirical evidence that this is true for a parent country's out-  In general, the interdependence across markets will depend on the economic proximity across the host markets, which can be measured by natural bilateral trade flows among the host markets. We hypothesize that a large amount of natural (or predicted) trade flows between two host countries indicates that consumers in one of the two markets could be served cheaply from the other one. In that case, a given parent country's FDI in one of the two host countries should substitute FDI in the other. The empirical analysis of bilateral FDI needs to account for third market influences that decline in economic proximity. Accordingly, it seems natural to apply recently developed methods for spatially dependent data. In this paper, we applied a spatial 25 HAC estimator of the variance-covariance matrix developed by Kelejian and Prucha (2007) for estimation and testing.
In our sample of bilateral outbound FDI stocks within Europe, we find strong evidence for the impact of regional trade agreements on FDI. Also, our results indicate that spatial dependence is present in the data. This leads to This is consistent with the prevalence of export-platform FDI, where foreign subsidiaries are located in host markets from which large consumer bases can be served cheaply.
For economic policy, our results confirm the strong link between trade and foreign direct investment. In particular, we find that trade policy as reflected in RTA has an impact not only on trade but also on FDI. Trade liberalization of a set of parent countries with some host markets leads to a relocation of FDI from other hosts into the liberalizing ones. This suggests that trade regionalism (i.e., liberalization with a subset of countries in the world econ-26 omy) for a given parent country exerts positive direct effects on FDI flows in some host markets and indirect negative ones on others. However, the indirect negative effects in percent, at least for trade liberalization in Europe, tend to be small as compared to the direct positive ones. These results are timely given the ongoing enlargement of regional trade agreements. Blomström, M., Kokko, A., 1997. Regional integration and foreign direct investment: A conceptual framework and three cases.

Descriptive statistics
Tables A.1-A.3 summarize the descriptive statistics of the untransformed variables and the correlation matrices of the within transformed variables under maximum row normalization (Table A.2) and row normalization (Table   A.3).

and A.3 -Correlation matrices <
The correlation matrix also contains once time-lagged exogenous variables which are used in the sensitivity analysis below.
5. Sensitivity analysis: adding once-lagged explanatory variables to the model Table A.4 summarizes our findings from a sensitivity analysis, in which we added once time-lagged exogenous variables to the model. The results indicate that many of the estimated coefficients are insignificant. The reasons for this are the high partial correlation coefficients among the contemporaneous and the once-lagged within transformed determinants (see Tables A.2 and A.3).
> Table A.4 -Adjustment dynamics < The estimation is based on 3373 observations. To calculate the SHAC-standard errors we use the Bartlett-window with cutoff 0.2 which implies that 75% of the observations get the non-zero weight by the Bartlett window. The Newey-West estimates of the standard errors use one time-lag. *** significant at 1%; ** significant at 5%; * significant at 10%.

NORMALIZED BY LARGEST ROW SUM
Model 2 Model 1 Std.
Std.  Note: GDP refers to real gross domestic product in US dollars (base year is 2000), SK is gross secondary school enrolment in percent, and DIST is the great circle distance between two countries' capitals.