2.1 Reciprocity of FDI flows

Conventional dyadic models of FDI flows typically imply reciprocity. The institutional and cultural distance literature suggests that FDI is more likely to flow between a pair of countries that are institutionally and culturally similar, share common languages and colonial ties, and are in an alliance relationship or tied by migrant networks (e.g., Eden and Miller, Reference Eden and Miller2004; Leblang, Reference Leblang2010; Li and Vashchilko, Reference Li and Vashchilko2010). Note that this type of reciprocity is based on covariates. Existing literature has not yet studied the reciprocity arising from the interdependence of the outcome variable itself. That is, FDI from country i to country j increases the probability of investment from country j to country i.

We argue that the reciprocity of FDI stems from the fact MNCs act as agents of transmitting information about their home countries, which is facilitated by the existence of global production networks; The information diffusion helps to reduce host-country firms’ transaction costs of investing in MNCs’ home countries, thereby increasing the likelihood of reciprocal flows of investment. Investing in a foreign country incurs a “liability of foreignness.” Foreign firms have disadvantages compared with indigenous firms because the former are unfamiliar with the business practices and institutional environments and face a legitimacy issue due to greater scrutiny from the public in the host country (Hymer, Reference Hymer1976; Zaheer, Reference Zaheer1995; Kostova and Zaheer, Reference Kostova and Zaheer1999). This unfamiliarity and legitimacy issue induces extra business costs that often deter foreign entry. Existing MNCs actually help diffuse information about business practices and institutional environments in their home country (Simmons and Elkins, Reference Simmons and Elkins2004; Kwok and Tadesse, Reference Kwok and Tadesse2006). This kind of information diffusion is particularly strong in the existence of global production networks. For example, it can happen within MNC's global production networks via vertical linkages to their upstream suppliers and downstream customers. More generally, information diffusion can occur through the spillover of knowledge and management know-how from MNCs to local firms. Consequently, local firms in the host country acquire more information about investment opportunities, business practices, government policies, and so on, thereby reducing information asymmetry and the liability of foreignness when investing in MNCs’ home country.Footnote ³ All else being equal, we therefore should expect that firms in country i are more likely to invest in country j if firms from country j hold a high stock of investments in country i.

An old literature has studied MNCs’ oligopolistic expansion strategy (Kindleberger, Reference Kindleberger1969; Hymer, Reference Hymer1976), which can also result in the reciprocity of FDI. MNCs arise from exploiting their firm-specific assets to overcome imperfections in arm's-length markets (Dunning, Reference Dunning1992; Caves, Reference Caves1996). These proprietary assets include, for example, advanced technology, brand names, product differentiation, and managerial and advertising skills, which possess substantial economies of scale. To make the most use of their firm-specific assets and best exploit economies of scale, MNCs actively seek to expand market shares. Yet, an MNC's successful expansion into a foreign market generates disruptive effects, threatening the market positions of local firms. To secure their competitive positions and obtain advantages stemming from large-scale operations, local firms have incentives to undertake a rivalrous expansion with highly differentiated products into the home market of the MNC (Veugelers, Reference Veugelers1995).Footnote ⁴ This rivalistic strategy leads to reciprocal flows of FDI (Graham, Reference Graham1978). We expect this kind of reciprocal FDI flows is more prevalent among developed countries because they are primarily market seeking and firms in developed countries have accumulated sufficient intangible assets that enable them to succeed in foreign markets.

Historically, global investment activities have been dominated by MNCs from developed countries and characterized by a pattern of two-way flows (Julius, Reference Julius1990, 22). Although this means reciprocity has been conditional on the level of development of the states in a dyad, the pattern has been changing. On the one hand, developing countries become increasingly popular investment destinations. In 2012, developing countries as a whole received more FDI than developed countries for the first time ever (UNCTAD, Reference UNCTAD2013). On the other hand, we have also witnessed a surge of FDI from, and increasingly between, developing countries since the beginning of the 21^st century. The percentage of outward FDI from developing and transition economies has grown from 8.78 percent in 2001 to 29.42 percent in 2017, with a peak in 2013 (approximately 35.54 percent of the world total). Yet, investing abroad incurs large fixed costs and firms need to overcome the disadvantages such as liability of foreignness they face when competing with indigenous firms in the host country. Therefore, only the most productive firms are able to engage in FDI activities (Helpman et al., Reference Helpman, Melitz and Yeaple2004). Given that MNCs from developing countries are still smaller and less competitive than their counterparts from developed countries, we expect that developing countries are less likely to reciprocate FDI. In our empirical analysis, we explicitly model this actor heterogeneity.

2.2 Transitivity/clustering of FDI flows

Transitivity, sometimes referred to as clustering, would manifest as dense triads of FDI emerging in the FDI network. We argue that the transitivity of investment activities arise from the expansion of global production networks and a safety net created by existing FDI linkages. One distinct feature of today's globalization is the increasing fragmentation and dispersion of production processes and the dramatic expansion of global supply chains (Baldwin, Reference Baldwin2011; UNCTAD, Reference UNCTAD2013). At the center of global production networks are MNCs, which coordinate global supply chains through complex networks of their foreign affiliates, subcontractors, or arm's-length suppliers (UNCTAD, Reference UNCTAD2013, xxii). These intertwined production networks give rise to the transitivity of FDI activities.

In a direct way, an MNC's establishment of foreign affiliates will be followed by investment of their upstream suppliers or downstream customers, who have also invested in the MNC's home country. This type of production connection will lead to multiple triadic closures of investment flows. Consider a case of three countries: A, B, and C. Suppose a firm from A invests in B as a supplier to a firm in B. If the firm in B establishes a foreign affiliate in C, the firm from A likely follows and makes an investment in C to serve the foreign affiliate, thereby leading to a triadic closure of investment flows.

Note that suppliers do not necessarily need to undertake an overseas investment to serve a leading firm. Alternatively, suppliers can choose to export intermediate goods to a leading firm. Yet, the aforementioned triadic closures of investment flows are not uncommon because leading firms tend to favor near suppliers due to an industry clustering effect or high transportation and trade costs (Carr et al., Reference Carr, Markusen and Maskus2001). For example, Volkswagen's investment in Skoda Auto in Czech Republic not only attracted other auto makers such as PSA Peugeot and Toyota, but also its international suppliers of parts and components to acquire local firms or build new factories; “As of 2002, there were 270 firms operating in the Czech Republic, representing 45 percent of the top 100 world suppliers of automotive parts and components” (Kaminski and Javorcik, Reference Kaminski, Javorcik and Broadman2005, 352). Similarly, Volkswagen's recent investment in Ningbo-Hangzhou Bay New Zone in China has brought in suppliers from South Korea, France, and the United States.Footnote ⁵ Airbus's establishment of its A320 assembly line in Tianjin, China has attracted investments from its global suppliers to the Tianjin Airport Economic Zone and contributed to an aviation industry cluster with an output estimated to exceed 100 billion Chinese yuan in 2020.Footnote ⁶ In India, investments by auto makers such as Toyota, Hyundai, and Ford also brought in their international component suppliers (Moran, Reference Moran2014, 23).

More generally, when two countries are tied to a third country via FDI, these FDI linkages create a safety net that protects MNCs from host governments’ opportunistic and predatory behavior and thus facilitates capital movements between these two countries, thereby resulting in a triadic closure of investment flows (i.e., the transitivity of FDI). Although direct asset expropriation has been rare since the late-1970s, political risk such as subtle policy changes remains a primary concern of investors (Graham et al., Reference Graham, Johnston and Kingsley2018), because footloose capital becomes an “obsolescent bargain” due to its ex post immobility (Vernon, Reference Vernon1971, Reference Vernon and Winchester1980). When two countries are connected to a third country via FDI, which typically involves the expansion of global production networks, it significantly constrains governments’ policy discretion. This is because the proper functioning of global production networks hinges crucially on the cooperation and coordination of the countries involved. For example, Johns and Wellhausen (Reference Johns and Wellhausen2016) show that host governments are less likely to expropriate foreign firms when they are closely connected to firms in host countries through supply chains. Dorussen and Ward (Reference Dorussen and Ward2010) demonstrate that countries are less likely to have conflicts with each other when they are more embedded in the trade networks. Similarly, Kim and Solingen (Reference Kim and Solingen2017) find that East Asian countries that are deeply integrated into global production networks are more likely to promote cooperation and peace between each other.

This risk-mitigating effect of the safety net is magnified when two countries are integrated into the same global production networks coordinated by leading MNCs in a third country. Consider an example that MNCs in country A offshore their production or other operating activities via FDI to countries B and C; local firms in countries B and C may also be integrated into these production networks as either upstream suppliers or downstream customers. In such a case, all three countries have strong incentives to ensure the well functioning of the network for economic benefits and thus are less likely to engage in opportunistic or predatory behavior, thereby providing a safety net for investors. Consequently, firms in countries B and C are also more likely to invest in each other, contributing to the transitiveness of FDI.Footnote ⁷ In general, when two countries are tightly linked to a third country through investment flows, FDI should be more likely to flow between these two countries due to shared economic interests and reduced political risk. Therefore, we expect that FDI has a high probability to flow among two countries that have strong investment ties to the same third country, resulting in the transitivity/clustering of investment activities.

It is important to note that our theory suggests that the reciprocity and transitivity of FDI is not a result of the global bilateral tax treaty (BTT) networks or the existence of tax havens.Footnote ⁸ Scholars have studied the implications of global BTT networks and tax havens for MNCs to create complex corporate structures for tax avoidance (e.g., Arel-Bundock, Reference Arel-Bundock2017b). For example, Google's “Double Irish Dutch Sandwich” corporate structure helped lower the company's effective tax rate to 2.4 percent on non-US income.Footnote ⁹ This corporate structure results in complex financial flows between the parent company and its foreign affiliates and among foreign affiliates themselves. Yet, a complex single-firm corporate structure or financial flows within the corporate structure do not lead to the reciprocity and transitivity of FDI. This is because investments from the parent company to its foreign affiliates are one-way flows and profits repatriation and interest, dividends or royalties payments within the corporate structure are not registered as FDI (see, Kerner, Reference Kerner2014).

3.1 Covariates

In the gravity model, we include the log product of the dyad's real GDP and logged Euclidean distance. Generally, higher GDP represents a larger market and therefore should be associated with more FDI, while an increase in geographic distance increases investment costs, decreasing investment flows. For the purpose of model convergence, the logged product of dyadic GDP has been estimated as one variable in the model, rather than being estimated separately. In addition, we include both origin and destination countries’ GDP per capita to roughly control for relative factor endowments.Footnote ¹⁴

Other economic controls include origin and destination countries’ trade openness (trade as % of GDP) and bilateral trade volumes between the origin and destination countries. Existing research has shown that FDI and trade are compliments (Markusen, Reference Markusen1995; Aizenman and Noy, Reference Aizenman and Noy2006). We expect that higher levels of trade openness and bilateral trade will be associated with higher levels of bilateral FDI. Trade openness data are from the World Bank's World Development Indicators and trade volume is from the OECD (2016).

There is a substantial amount of research that explores the relationship between democratic institutions and FDI inflows; yet empirical results to date remain inconclusive (see, e.g., Jensen, Reference Jensen2003; Jakobsen and De Soysa, Reference Jakobsen and De Soysa2006; Arel-Bundock, Reference Arel-Bundock2017a; Li et al., Reference Li, Owen and Mitchell2018; Wright and Zhu, Reference Wright and Zhu2018). We include standard polity scores as a measure of a country's level of democracy (Marshall and Jaggers, Reference Marshall and Jaggers2010). We also include three international institutions variables: BITs, PTAs, and OECD membership. BITs are included as a binary variable that is one if the pair have a stand-alone BIT or are party to a preferential trade agreement that also covers investment policy. These treaties should be positively associated with FDI levels as they should effectively remove barriers to investment and provide commitment to liberal economic policies (e.g., Büthe and Milner, Reference Büthe and Milner2008; Osnago et al., Reference Osnago, Rocha and Ruta2016). PTAs vary significantly in depth with some requiring nearly full liberalization of trade barriers while others are superficial political signals. Therefore we make use a recent PTA depth variable that uses latent trait analysis with 48 different dichotomous variables regarding topics covered in PTAs.Footnote ¹⁵ Signing a PTA represents a commitment to liberal markets that investors would favor and therefore would be associated with increased FDI inflows (Büthe and Milner, Reference Büthe and Milner2008, Reference Büthe and Milner2014). OECD membership is included as a binary node-level homophily variable that is one when both countries are either OECD members or both are not OECD members. This variable adjusts for OECD-based homophily and is expected to be positive.

In addition, we include two sets of international agreement variables. The first is a binary variable for a combination of military alliance treaties that are not defense treaties. The second is a defense treaty. Both are from Gibler (Reference Gibler2009). We expect these variables to be positively associated with FDI inflows, particularly defense treaties since this indicates political cooperation and low political risk (Li and Vashchilko, Reference Li and Vashchilko2010).Footnote ¹⁶

3.2 Model and specification: the count ERGM

Like other forms of the ERGM, the count ERGM is a statistical model that operates on a complete network. To specify the count ERGM, the researcher selects two types of network statistics—those that relate tie values to observed covariates (i.e., covariate effects), and those that relate the ties to each other via higher-order network structure (i.e., network effects). The ERGM family of models is innovative in that both of these statistic types—covariate effects and network effects—can be included in the same model. Including the network effects helps to accurately identify the true covariate effects (Metz et al., Reference Metz, Leifeld and Ingold2019). Under Krivitsky's (Reference Krivitsky2012) count ERGM, the probability of the observed n × n network adjacency matrix y is

(1)$$ {\rm Pr}_{{\boldsymbol \theta}; {h}; {g}}( {\boldsymbol Y} = {\boldsymbol y} ) = {h( {\boldsymbol y}) {\rm exp}( {\boldsymbol \theta} \cdot {g} ( {\boldsymbol y}) ) \over {\kappa}_{{h}, {\rm g}}( {\boldsymbol \theta}) },\; $$

where g(y) is the vector of network statistics used to specify the model, θ is the vector of parameters that describes how those statistic values relate to the probability of observing the network, h(y) is a reference function defined on the support of y and selected to affect the shape of the baseline distribution of dyadic data (e.g., Poisson reference measure), and κ_h,g(θ) is the normalizing constant.

3.2.1 Specification

In the models we specify, we use statistics that model the shape of the individual edge distributions (i.e., the shapes of directed dyadic FDI flows), model the dependencies we have described above, and account for the effects of exogenous covariates. Network statistics in an ERGM model local dependencies, just like conventional covariates, but are expressed at the network-level because that is how they are incorporated into the ERGM probability distribution (Desmarais and Cranmer, Reference Desmarais and Cranmer2012). Analogous to selecting covariates to include in a regression model, ERGM software packages present several options for adding network statistics to the model. The statistics we use to account for the individual edge distribution include (1) the sum of edge values, which models the average edge value; (2) the sum of square-root values of edges, which accounts for dispersion; and (3) the number of non-zero edges, which accounts for the prevalence of zeros in the edge values.

We include two statistics to model the dependencies that correspond to our hypotheses. First,

(2)$${\rm Reciprocity}\,\colon\, {\boldsymbol g}( {\boldsymbol y}) = \sum_{( { i, j}) {\in} {\opf Y}}\min( {\boldsymbol y}_{ i, j},\; {\rm y}_{ j, i}) ,\; $$

in which we add up the lowest edge value within each dyad. If edges are reciprocated, this statistic will increase due to the co-occurrence of large edge values within the same dyad. To evaluate our hypothesis regarding the higher degree of reciprocity among developing countries, we fit two separate reciprocity terms—one that applies to dyads in which both states are OECD members, and one that applies to every other dyad (i.e., both non-OECD, and mixed dyads that include an OECD and non-OECD state).Footnote ¹⁷ Second,

(3)$${\rm Transitive\, Weights}\,\colon\, {\bf g( y) } = \sum_{( i, j) {\in} {\opf Y}}\min\bigg({\bf y}_{i, j},\; \max\limits_{k{\in}N}\Big(\min( {\boldsymbol y}_{i, k},\; {\boldsymbol y}_{ k, j}) \Big)\bigg),\; $$

which accounts for the degree to which edge (i, j) co-occurs with pairs of large edge values with which edge (i, j) forms a transitive (i.e., non-cyclical) triad with weighted, directed two-paths going from nodes i to k to j. Exogenous covariates are accounted for with statistics that measure the degree to which large covariate values co-occur with large edge values. First,

(4)$${\rm Dyadic\, Covariate}\,\colon\, \bi{g( y,\; x) } = \sum_{( i, j) } \bi{y}_{i, j}{\rm x}_{i, j},\; $$

measures this co-occurrence at the level of the directed dyad, in which there is a dyadic observation of the covariate corresponding to each potential FDI flow. Second, there are two statistics that account for node (i.e., country) level covariates. Each statistic takes the product of the node's covariate value and a sum of the edge values in which the node is involved. The first, “Sender Covariate,” uses the sum over the edges that the node sends. The second, “Receiver Covariate,” uses the sum over the edges that the node receives.

We use the count ERGM implementation of Krivitsky's (Reference Krivitsky2012) made available in the ergm.count (Krivitsky, Reference Krivitsky2016) package in R statistical software. The normalizing constant in the likelihood function for the count ERGM (i.e., the denominator in Equation 1) is given by

(5)$$\bf{\kappa}_{\boldsymbol h, \bf{g}}( \bf{\theta}) = \sum_{\bf{y} \in \bf{{\cal Y} } } {\boldsymbol h}( {\boldsymbol y}) {\rm exp}( \boldsymbol {\theta} \cdot \bi{g} ( \bi{y}) ) ,\; $$

where $\bf {{\cal Y} }$ is the set of all possible count network configurations, and h(y) = ∏_i,j(y _i,j!)⁻¹ is the Poisson reference measure, which assures that (1) the normalizing constant is a convergent sum, and (2) produces an edge-wise conditional Poisson distribution if there are no dependence terms in the model. As with the binary ERGM, the normalizing constant in the count ERGM is computationally intractable, and must be approximated. Estimates of θ are calculated using Monte Carlo maximum likelihood (MC-MLE).

We estimate a separate model for each year from 2002 to 2012. We have two main reasons for presenting year-by-year estimates as our main results. First, since analyzing dyadic data essentially squares the size of the data when compared to the monadic level, we have enough data to identify a separate set of parameter values in each year. Second, recent international relations applications have called into question the appropriateness of pooling over long time periods since there may be considerable historical heterogeneity in the parameter values, and have estimated separate models for each year or time period (see, e.g., Ward and Hoff, Reference Ward and Hoff2007; Cranmer et al., Reference Cranmer, Heinrich and Desmarais2014). In the results we present below, we see that many of the parameters vary considerably over time. In particular, several parameters exhibit significant shifts in magnitude and statistical significance beginning in 2008—a pattern that is likely attributable to the Great Recession. We present time-pooled results in Supplementary Appendix C, and find that our results are robust.