Multinational ownership, intellectual property rights, and knowledge diffusion from FDI

Transcription

1 NiCE Working Paper November 2008 Multinational ownership, intellectual property rights, an knowlege iffusion from FDI Roger Smeets Albert e Vaal Nijmegen Center for Economics (NiCE) Institute for Management Research Rabou University Nijmegen P.O. Box 9108, 6500 HK Nijmegen, The Netherlans

2 Abstract In this paper we exten the vertical linkages moel by Markusen an Venables (1999) to inclue (a) i ering egrees of multinational (MNE) ownership in their foreign a liates an (b) knowlege i usion, in aition to eman an supply linkages. We investigate the intra- an interinustry e ects of changes in MNE ownership on local rms prouctivity via eman linkages, price e ects an knowlege i usion. Moreover, we also consier the meiating in uence of national intellectual property rights protection (IPP). Given the ambiguous preictions of our moel, we also investigate these issues empirically in a panel of 1222 large rms sprea out over 20 countries an 18 manufacuring inustries uring the perio : We n that in countries with low IPP, the occurence of intra-inustry prouctivity e ects is conitional on the cost structure of local rms. Moreover, inter-inustry prouctivity e ects are largely absent. Conversely, in countries with high IPP, both intra-inustry an inter-inustry prouctivity e ects are high. Also, the relationship beween prouctivity e ects an MNE ownership varies both within an between inustries, as well as between conitional an unconitional prouctivity e ects. We interpret this empirical evience as a con rmation of our theoretical conjecture that intra-inustry knowlege i usion is ominate by unintene spillovers, whereas inter-inustry knowlege i usion is ominate by intene knowlege transfers. 1

3 1 Introuction In an attempt to better isentangle the conitions uner which Foreign Direct Investment (FDI) inuces knowlege spillovers, acaemic research has increasingly taken into account the heterogeneity of multinationals (MNEs) an their foreign subsiiaries (Feinberg an Keane, 2005; Smeets, 2008). Whereas FDI use to be treate as a rather bulky an homogeneous concept (Lipsey, 2002), scholars have starte to acknowlege the heterogeneity of MNEs in inter alia investment motives (Girma, 2005; Dri el an Love, 2007), market orientation (Girma et al. 2008) an country of origin (Javorcik et al., 2004; Girma an Wakelin, 2007), an the subsequent consequences for host-country knowlege spillovers. A particularly promising stran of research has consiere i erences in MNE ownership over foreign a liates as a etermining factor of knowlege spillovers (Blomström an Sjöholm, 1999; Dimelis an Louri, 2002; Javorcik, 2004; Javorcik an Spatareanu, 2008). Internalization theory suggests that increase MNE ownership over a foreign a liate inuces the parent to transfer more proprietary knowlege or technology abroa, thus increasing the potential for knowlege i usion (Rugman, 1981; Hennart, 1982; Davies, 1992). Moreover, stuies on MNE input sourcing suggest that increase MNE ownership has consequences for the extent of local input sourcing, thus a ecting the extent of backwar linkages (Tavares an Young, 2006; Javorcik, 2008). Empirical stuies usually istinguish minority from majority ownership (Blomström an Sjöholm, 1999; Dimelis an Louri, 2004), or share ownership from fully owne subsiiaries (Javorcik, 2004; Javorcik an Spatareanu, 2008) an inee n that the istinction matters. However, none of these stuies consiers the e ect of MNE ownership as a continuous variable, which veils a lot of the potential variation. Moreover, although some theoretical stuies investigate the relationship between intra- rm knowlege transfer an MNE ownership (Müller an Schnitzer, 2006), theoretical contributions on the relationship between MNE ownership an intra an inter-inustry knowlege i usion are largely absent. This paper rst picks up on the latter observation: We introuce share ownership between a MNE an a local (host-country) partner in the foreign subsiiary as a variable of interest in a theoretical moel by Markusen an Venables (1999), an then consier its host-country intra an inter-inustry e ects. Speci cally, in aition to consiering only pecuniary externalities, as is common in most theoretical moels, we also consier actual knowlege i usion. In oing so, the analysis explicitly consiers two forms of knowlege i usion: First, knowlege spillovers, which are unintene knowlege ows (i.e. externalities) from the MNE to its host-country environment. Secon, knowlege transfers, which are intene (an internalize) ows of knowlege from the MNE to its host-country environment. A nal contribution of the paper is that it also consiers institutional heterogeneity - notably intellectual property rights protection (IPP) - an how this interacts with the two types of knowlege i usion just mentione. Our theoretical results emonstrate the opposing in uences of pecuniary 2

4 e ects, irect an inirect eman e ects, an knowlege i usion on omestic rms, following from an increase in MNE ownership in foreign a liates. Nonetheless, we are able to erive some (conitional) unambiguous preictions: We n that forwar or ownstream host-country e ects following an increase in MNE ownership are generally positive, provie that there is su cient ownstream competition. Backwar or upstream e ects are also positive in countries with high IPP, provie that inter alia ownstream eman elasticities an local input shares are su ciently high. The intra-inustry e ects are generally positive in low-ipp countries, provie that the share of xe costs in total costs of omestic rms is su ciently high. We then take these theoretical preictions to the ata, by employing a rmlevel panel ataset, containing 1222 large omestic rms an 351 foreign subsiiaries with varying egrees of MNE ownership, active in 20 countries an 18 inustries uring the perio Our theoretical nings on the interinustry e ects are largely con rme by the ata. The empirical results on the intra-inustry e ects are not entirely in line with the theoretical expectations, which we argue might be ue to some assumptions in the moel. Generally speaking, the empirical results suggest that high-ipp countries are better able to reap the bene ts of MNE investment than low-ipp countries. The empirical part of the paper also provies two methoological avantages over some of the earlier stuies alreay unertaken in this area (cf. Blomström an Sjöholm, 1999; Dimelis an Louri, 2002; Javorcik, 2004; Javorcik an Spatareanu, 2008). First, unlike earlier stuies, we utilize a cross-country sample which allows us to investigate how institutional heterogeneity (such as i erences in IPP regimes) interacts with the relationship between MNE ownership an host-country prouctivity e ects. Secon, in stea of consiering ichotomous or iscrete i erences in MNE ownership (e.g. minority versus majority, or share ownership versus full ownership), we treat MNE ownership as a continuous variable in the empirical part as well. Given the ex ante theoretical ambiguity of the relationship between MNE ownership an knowlege i usion, next to the usual parametric regression techniques we also employ semi-parametric regression techniques which allows us to refrain from specifying a speci c functional form regaring the relationships of interest. The rest of this paper is structure as follows: Section 2 evelops a theoretical moel, base on Markusen an Venables (1999) an extens it with the theoretical elements mentione above. Section 3 analyzes the within an between inustry-e ects of changes in MNE ownership on local rms, an how these effects epen on the extent of IPP. Section 4 outlines the empirical methoology an gives an overview of the ata use in this paper. Section 5 presents the estimation results of the empirical moel. Finally, Section 6 conclues. 2 The moel Before iscussing the setup of the moel, it is instructive to consier Figure 1 below, which presents a schematic representation of the theoretical moel. A 3

5 MNE sets up a share foreign subsiiary in sector k in the host economy to prouce for an sell on the local market. As such, it competes with local rms that are also active in sector k, but at the same time it also spills over knowlege to these rms. These are intra-inustry or horizontal knowlege spillovers. The inustry the MNE invests in may be a ownstream inustry receiving inputs from local rms in j as inicate by situation A or an upstream inustry elivering inputs to rms in inustry l. If situation A is at han, local rms active in sector j supply the foreign subsiiary an local rms in sector k with intermeiates, but at the same time also receive knowlege transfer from the foreign subsiiary, e.g. through supplier assistance (Javorcik, 2008). This type of knowlege transfer is calle backwar or upstream knowlege transfer. If situation B is relevant, the foreign subsiiary an local rms in sector k may also function as input suppliers themselves, selling intermeiates to local rms in sector l. Simultaneously, these local sector l rms may receive knowlege transfer from the foreign subsiiary, e.g. in the form of increase input quality (Javorcik, 2008). Whichever situation occurs, these types of knowlege transfer are inter-inustry or vertical in nature. In aition to vertical knowlege transfer, changes in the eman an supply of goos along the input an output linkages will cause backwar an forwar eman e ects, leaing to pecuniary spillovers. << INSERT FIGURE 1 ABOUT HERE >> In what follows we will rst focus on part A of Figure 1. That is, we will rst consier the situation in which the MNE has a share subsiiary in the ownstream sector (k) of the host economy an receives supplies from the upstream sector (j). After having erive the moel for this setup, we will also inicate how the moel changes when consiering part B, where the foreign subsiiary is active in the upstream sector (k), supplying local rms in the ownstream sector (l). In our moel there are two types of rms: Multinationals (m) an national rms (n), the latter of which can be further classi e as local partners (lp), ownstream rms () an upstream rms (u). We assume that MNEs require a local partner to set up a foreign subsiiary in the host country: The resulting share subsiiary can be thought of as an International Joint Venture ( ). 1 This competes with the ownstream rm, an both of them are supplie by the upstream rm u. The theoretical moel below buils on an extens Markusen an Venables (1999). These authors evelop a multi-sector partial equilibrium moel, where they analyze the e ect of MNE entry in a ownstream inustry on the number of 1 A couple of remarks apply here: First, note that we assume that the MNE nees a local partner, i.e. we o not moel the ecision between a green el versus a share subsiiary nor the search process for a suitable partner. Secon, even though the share subsiiary may be thought of as an, we assume that the national rm is completely absorbe in the partnership an oes not have any remaining operations of its own. From that perspective, the partnership may have more resemblance to a partial acquisition. Thir, we assume that there is always a su cient supply of local partners. 4

6 local rms active in upstream an ownstream inustries. The e ects of MNE entry work via competition e ects an eman linkages (leaing to pecuniary externalities). Our setup resembles theirs in a number of ways: We also utilize a two inustry setup, in which each inustry is characterize by Dixit-Stiglitz monopolistic competition. Further, we also look at pecuniary externalities via eman linkages between the upstream an ownstream inustries. Yet our moel also i ers from theirs in three important aspects: First, we introuce share ownership between the MNE an a local partner in the foreign subsiiary, as investigating the e ect of a change in MNE ownership on the prouctivity of local rms is the primary focus of this paper. Secon, next to pecuniary spillovers, we also introuce irect an explicit knowlege i usion. Moreover, we isentangle these knowlege i usion e ects into knowlege spillovers (horizontal) an knowlege transfers (vertical), an consier their contingency on IPP protection. Finally, we o not consier the e ect of MNE entrance or ownership on the entry or exit of local rms, by keeping the number of rms constant when taking total erivatives (cf. Section 3). Once again recall that we rst consier part A of Figure 1, where s are active in the ownstream inustry (together with local ownstream rms ) an are supplie by local upstream rms u. We moel the price inex of the inputs prouce by local upstream rms in CES fashion an enote it by: P U = n u p 1 1=(1 ) (1) u where n u are the number of local upstream rms, p u are iniviual prices of upstream inputs an > 1 is the elasticity of substitution between any two input varieties. Suppose for the moment that total eman for inputs from the ownstream sector is given by I. Then, multiplying P U by I gives total costs of input supply, or equivalently, total expenitures on inputs. Hence, we can apply Shephar s lemma to erive eman for iniviual inputs x u : x u = p u IP U (2) In the ownstream sector we have a similar inustry structure, but here both national rms an s are active. Hence, the price inex in the ownstream sector is given by: P D = (n p 1 " + n p 1 " )1=(1 ") where n (n ) is the number of local rms ( s) active in the ownstream sector, p (p ) are the prices these rms charge, an " > 1 is the elasticity of substitution between any two varieties. The volume of total consumer eman for these ownstream proucts is given by Y an total expeniture on D ownstream goos is given by Y P where is the elasticity of eman with respect to the price inex P D. Similar to Markusen an Venables (1999), we assume that " >.> 1. Again applying Shephar s lemma we obtain iniviual emans: x = p " Y P " D (3) x = p " Y P " D 5

7 First consier the pro t function of the which is given by: = p x (F + x ) [P U + (1 )w] (4) where p enotes price, x enotes output, F are xe costs, are marginal prouction costs, w is the wage rate of labor, an is the share of inputs source from the upstream sector (0 1). Note that the amount of inputs source from the upstream sector epens on the amount of xe costs an variable costs. The remaining share (1 ) is spent on labor as an aitional prouction factor. As mentione, the is a partnership between a MNE (m) an a local partner (lp). We assume that the contribution of both rms in terms of technology an knowlege to the is proportional to their ownership shares in the, which is given by for the MNE an (1 ) for the local partner. These contributions translate into the xe an marginal prouction costs of the an are moelle as follows: F = F m + (1 )F n (5) = m + (1 ) n where we assume F lp = F = F u F n, i.e. xe costs of all national rms are equal, regarless of their type, an similarly for. In line with earlier literature (Blomström an Sjöholm, 1999), as well as with the rm characteristics in our own sample (see Section 5), we assume that F m < F n an m < n i.e. the MNE is more prouctive than a national rm, both in terms of xe costs as well as marginal costs. Hence, the larger the ownership share of the MNE in the, the lower xe an marginal costs will be, which is in line with the literature on internalization or transaction costs an technology transfer (Davies, 1992). A key issue of this paper is the nature an extent of knowlege i usion from the to the national rms. As we alreay explaine, we make an explicit istinction between unintene knowlege spillovers on the one han, an intene knowlege transfer on the other. This istinction is especially important in the present context, since we conjecture that the type of knowlege i usion is contingent on the irection of i usion, i.e. horizontal or vertical. Speci cally, we argue that knowlege spillovers from the are most likely to ow horizontally, i.e. to ownstream rms active in the same sector, for the has nothing to gain from intentionally transferring knowlege or technology to its competitors. Moreover, since these rms are active in the same sector, their absorptive capacity can be expecte to be relatively high. Intentional knowlege transfers on the other han, are more likely to ow vertically, i.e. from the to local upstream rms u (in situation A of Figure 1), since the will bene t from this by increase quality or ecrease prices of inputs. Inee, there exists ample evience of MNEs that assist their suppliers in terms of technology transfer, or transfer of best practices or quality stanars (Javorcik 6

8 an Spatareanu, 2005; Javorcik, 2008). 2 In the context of knowlege i usion, the extent of IPP also becomes relevant (Branstetter et al., 2006) since the purpose of IPP is to reuce knowlege spillovers. As a consequence we may expect opposite e ects of IPP on (horizontal) knowlege spillovers on the one han, an (vertical) knowlege transfer on the other han: If IPP functions properly, horizontal knowlege spillovers shoul be reuce. At the same time however, ue to the ecrease risk of expropriation of knowlege, this increases the incentives for the to (vertically) transfer knowlege. Hence, upstream knowlege transfer shoul increase with IPP. 3 As we have assume that MNE knowlege transfer to the takes e ect through xe an marginal costs, it is only natural to assume that knowlege i usion from the to ownstream an upstream rms will also a ect their xe an marginal cost structures. Hence, for local ownstream rms, we moel xe an marginal costs after spillovers as: F S = F + (1 )F (6) S = + (1 ) where is a parameter capturing the strength of Intellectual Property Rights protection (IPP), with = 1 enoting perfect protection an = 0 no protection whatsoever. Hence, spillovers are maximize when = 0, implying that the xe an marginal cost structures of s can be copie perfectly. For intentional knowlege transfers from the to local upstream rms we then have: F T u = (1 )F u + F (7) T u = (1 ) u + Note that because knowlege transfer is intentional (as oppose to spillovers) the is more willing to transfer its technology as the extent of IPP increases ( increases), since the risk of expropriation is very small in that case (Branstetter et al., 2006). The local upstream rm has the following formulation for pro ts: u = p u x u (F T u + T u x u )w (8) We can erive the equilibrium price for the upstream rm by substituting equilibrium eman (2) into (8) an maximize pro ts, which yiels: p u = T u w ( 1) 2 We o not consier explicit learning within the by any of the two parties involve (for an analysis of this type, see Müller an Schnitzer, 2006). 3 Apart from the theoretical relevance of introucing IPP in this manner, its opposite e ects on knowlege spillovers an transfers also allow us to test our hypothesize i erence between horizontal an vertical knowlege i usion empirically. 7

9 It irectly follows from this expression that MNEs bene t from technology transfer to upstream rms, since this ecreases T u an hence ecreases input prices p u. Local ownstream rms have the following pro t function: = p x (F S + S x )(P U + (1 )w) (9) the interpretation of which is similar to that of the. 4 The equilibrium pricing conition is foun by substituting x from (3) into (9) an maximizing pro ts: p = "S (P U + (1 (" 1) )w) Note that on top of the knowlege spillovers through S, the backwar eman linkage from MNEs to upstream rms poses an aitional bene t to the local ownstream rm as it serves to ecrease P U as well, which constitutes an (inirect) forwar linkage. Finally, for the we obtain a similar pricing conition: p = " (P U + (1 (" 1) )w) We can now close the moel by also writing own erive eman for the upstream rm s proucts, which is generate by the input eman from the an the omestic rm in the ownstream sector: 5 I = n (F + x ) + n (F S + S x ) (10) So far, we have only consiere part A of Figure 1, i.e. the situation in which the is active in the ownstream sector generating horizontal intrainustry e ects as well as upstream or backwar e ects through inter-inustry linkages. In orer to analyze ownstream or forwar linkages, we also consier the situation in which the is active in the upstream inustry (together with local rms) an supplying local rms in the ownstream inustry. That is, part B of Figure 1. 6 Because the moel remains largely the same, except 4 Note that we assume (unlike Markusen an Venables, 1999) that = = : Although it has been argue that MNEs (or s) will potentially source less of their inputs in the host-country, we have no way of istinguishing between an in the empirical part of the paper, so that we prefer the current speci cation. However, we will come back to the implie relationship between an MNE ownership when iscussing the empirical results later on. 5 Coming back to our earlier remark, we again note that we refrain from eriving free entry (i.e. zero pro t) conitions, but instea assume that these are ful lle in both sectors. A potential problem in this case is that the cost structure of the two rm types in the ownstream sector ( s an s) i er. Speci cally, given that s are more e cient than s, imposing a zero-pro t conition for s woul imply positive pro ts for s. In orer to prevent this situation from ocurring, we assume that any resulting positive pro ts from s are absorbe by ae co-orination costs between the MNE an its local partner. 6 We alreay note above that in the moel setup iscusse so far, we o have inirect forwar linkages to the ownstream local rms which are contingent on the upstream linkage, 8

10 for the fact that the switches inustries, we will not fully write it own here (Appenix A). However, note that in this case it is the upstream rm that bene ts from knowlege spillovers, whereas the ownstream rm bene ts from knowlege transfer. This also implies that the moerating e ects of IPP change accoringly. In the next section we will analyze the comparative static e ects of a change in MNE ownership in the () on the pro ts of local rms for both situations A an B. 3 Intra an inter-inustry e ects of MNE ownership 3.1 s in the ownstream sector Since our main interest in this paper concerns the e ects of MNE ownership in the () on local rms through eman linkages, competition e ects an knowlege i usion, we investigate the e ect of on local rms pro ts. In orer to o this, we compute total erivatives with respect to while assuming that all other variables remain unchange. First consier the e ect of MNE ownership in the ownstream inustry on upstream rms pro ts: u = p1 u P U BL {z} 1?0 + p u P E 1 {z} <0 + KT 1 {z} >0 (11) where BL 1, P E 1 an KT 1 are a backwar linkage e ect, a price e ect an a knowlege transfer e ect respectively, the full expressions of which are given in Appenix B1. The knowlege transfer e ect KT 1 is straightforwar: An increase in MNE ownership in the increases explicit knowlege transfer to the upstream rm by ecreasing xe an variable costs, increasing upstream rms pro ts. Moreover, the larger the IPP (i.e the larger ), the larger is this positive e ect. The negative upstream price e ect P E 1 is ue to our assumption of homogeneity of rms an their interrelationships, so that all upstream rms are a ecte by an increase in in the same way. Speci cally, the ecrease in T u following an increase in ecreases upstream prices p u. This e ectively reuces the price inex in the upstream inustry P U, i.e. it epresses per rm revenue in this sector. This e ect is stronger the larger is ue to increase knowlege transfer. The e ect of through the backwar eman linkage (BL 1 ) has three components in u = (see Appenix). First, there is a negative inirect knowlege spillover e ect, which occurs because of the increase in knowlege spillovers to the local ownstream rm as a result of an increase in, making ownstream since they will be a ecte by changes in P U inuce by changes in MNE ownership in the (). However, in the empirical section, we will also investigate the irect forwar linkages, i.e. the linkage e ects of an irectly supplying local rms, so that we also have to consier this case theoretically. 9

11 rms more e cient. This implies less eman for x u since less inputs are neee to prouce the same output. Also note that the negative e ect of knowlege spillovers is moerate by the extent of IPP: The larger, the smaller knowlege spillovers to the local ownstream rm an hence, the smaller its negative in uence on eman for intermeiate inputs. This as to the positive irect e ect of through KT 1. Secon, there are positive ownstream eman e ects, inuce by the change in eman for ownstream rm proucts after an increase in. Since xe an marginal costs of ownstream rms are reuce, as well as the fact that input prices P U go own, prices for ownstream proucts fall, inucing an increase in eman for ownstream proucts an accoringly also for upstream inputs. The impact of on this e ect is twofol: On the one han, an increase in ecreases knowlege spillovers to local ownstream rms, thus limiting the price ecrease of these rms an limiting the increase in erive input eman. On the other han, an increase in raises knowlege transfer to the upstream rm, lowering input prices an ownstream prices, thus increasing erive eman for inputs again. However, this latter e ect is a secon orer e ect, so that in this case, IPP will most likely exert a negative e ect on u. The thir e ect which takes place through the backwar eman linkage (BL 1 ) is a ownstream price e ect. As we will see below as well (when analyzing =) an increase in ecreases iniviual prices of all ownstream rms, an thereby also the price inex P D. That is to say, an increase in eventually ecreases per rm revenue in the ownstream sector. This in turn has a negative impact on erive eman for upstream inputs an accoringly on upstream pro ts. We now turn to the national rm in the ownstream inustry (i.e. the competitor of the ). Recall that two clear i erences with the upstream rm are that (i) the ownstream rm is not vertically linke with the an (ii) knowlege i usion occurs through knowlege spillovers rather than knowlege transfer. Computing the total erivative of with respect to yiels: = (" )p " x P E 2 P D {z} <0 + KS {z} 1 >0 + IDL 1 {z } >0 (12) where P E 2, KS 1 an IDL 1 are a price e ect, a knowlege spillover e ect, an an inirect eman linkage e ect respectively (the explicit expressions are relegate to Appenix B2). First, the price e ect (P E 2 ) works through the price inex P D. Due to the increase in, cost structures improve because of increase intra- rm knowlege transfer, making s more competitive. Moreover, since the increase in also increases horizontal knowlege spillovers, contingent on the lack of IPP (1 ), each iniviual national rm in the ownstream inustry is confronte with a ecrease in P D, a ecrease in per rm revenue, an hence a ecrease in pro ts. Secon, there is the irect knowlege spillover e ect (KS 1 ), occurring through the xe an marginal cost structure an again contingent on the absence of 10

12 IPP protection (1 ). This e ect is obviously positive. Finally, the ownstream national rm also pro ts from the vertical linkage between the s an the upstream rms, albeit in an inirect way via P U (IDL 1 ). Inee, since backwar knowlege transfer from the to the upstream rm increases with (see above), national ownstream rms are confronte with lower input prices P U. The extent of this positive inirect linkage is contingent on the input share as well as on the extent of IPP. Regaring the latter, this poses a counter-acting force to the irect knowlege spillovers from the to the ownstream rm: To bene t more from these spillovers, the ownstream rm requires a lower ( rst-orer), but to bene t from lower input prices, it requires a higher (secon-orer). 3.2 s in the upstream sector Now we will consier the situation in which the s (an local rms) are active in the upstream sector, hence supplying local rms in the ownstream sector (situation B in Figure 1). The analysis is similar to the one before, but recall that in this situation the local upstream rms bene t from knowlege i usion through knowlege spillovers, whereas the ownstream rms receive MNE knowlege through knowlege transfer. First consier the e ect of an increase in on local upstream rms pro ts (the explicit formulations are relegate to Appenix B3): u = p u (IDL 2 {z } <0 + P E {z} 3 ) + KS {z} 2 <0 >0 As before, there are three e ects: An inirect eman linkage e ect (IDL 2 ), a price e ect (P E 3 ) an a knowlege spillover e ect (KS 2 ). First, the inirect eman linkage takes e ect as a result of an increase in knowlege transfer from the to the local ownstream rms. As they become more e cient, the erive input eman ecreases, lowering upstream rms pro ts. Moreover, the higher IPP (i.e. the higher ), the more knowlege is transferre ownstream an the larger the negative e ect on u. Secon, as similar as before, the price e ect occurs because knowlege spillovers to local upstream rms an knowlege transfer by the MNE to the a ect all rms in the upstream sector simultaneously. This lowers the price inex P U an thus also per rm revenue. Moreover, the extent to which knowlege spillovers a to this e ect is larger the lower. Thir, the knowlege spillover e ect obviously increases upstream pro ts, an this e ect becomes stronger the lower is. Finally, for local ownstream rms we now have (the explicit formulations are relegate to Appenix B4): (13) = P E 4 {z} <0 + KT {z} 2 >0 + F L 1 {z} >0 (14) We can istinguish a price e ect (P E 4 ), a knowlege transfer e ect (KT 2 ) an a forwar linkage e ect (F L 1 ). The negative price e ect occurs because all 11

13 ownstream rms are similarly a ecte by knowlege transfer an ecrease input prices. That is, they all become more prouctive an charge lower prices, thus ecreasing per rm revenue. Note that has an ambiguous e ect on this mechanism: On the one han, it magni es the negative e ect through increase knowlege transfer, but on the other han it reuces it through ecrease knowlege spillovers to upstream rms (an hence, a smaller ecrease in input prices). The knowlege transfer e ect obviously increases ownstream rm pro ts, an the more so the higher is. Finally, the forwar linkage e ect occurs because knowlege spillovers from the to local upstream rms, as well as knowlege transfer from the MNE to the, ecrease input prices for ownstream rms, thus increasing their pro ts. In this case, an increase in has an unambiguously negative e ect, as it serves to ecrease knowlege spillovers. 3.3 Signing the e ects All total erivatives erive contain e ects that are opposite in sign. Moreover, as can be seen from the expressions in the Appenix, eriving conitions uner which their sign is unambiguous is not straightforwar for most of these erivatives. A lot of this ambiguity is cause by the often opposing e ects of IPP (). Inee, as it turns out, many of the expressions simplify substantially when consiering the extreme cases, i.e. when = 0 or = 1. In Table 1 below we summarize the signs of the total erivatives, also consiering the cases in which = f0; 1g. The erivations can be foun in the Appenix. << INSERT TABLE 1 ABOUT HERE >> First consier the e ects through backwar linkages, i.e. u = with MNEs ownstream. When is variable or when = 0, the e ect of a change in on upstream pro ts is ineterminate in (11). However, given that the conitions in Table 1 are met, the e ect is unambiguously positive for = 1. Obviously, = 1 inicates perfect IPP an upstream knowlege transfer by the MNE is at its maximum, ceteris paribus maximizing the positive e ect of KT 1 in (11). Moreover, the conition implies that the positive e ect is more likely (i) the smaller are total variable costs relative tot total xe costs in the ownstream inustry (the LHS of the conition), (ii) the larger is an (iii) the larger is. The latter e ect is cause by the fact that a higher - implying a more pricesensitive ownstream eman - translates the ownstream price ecrease (ue to a ecrease in P U following increase knowlege transfer) into higher ownstream eman, an thus also higher erive eman for upstream intermeiates. This e ect in turn is larger, the larger is the intermeiate input share of ownstream rms. The rst e ect occurs through the backwar eman linkage BL 1 : Since ownstream xe costs are only a ecte through knowlege spillovers, whereas marginal costs both through knowlege spillovers an price e ects, the total negative e ect on upstream rms of these combine e ects will be lower, the smaller are variable costs relative to xe costs. Next consier the e ects through forwar linkages, i.e. = with MNEs upstream. The table shows that in all cases, " > 2 is a su cient conition for this 12

14 erivative to be positive. The reason for this is that although the negative price e ect P E 4 in (14) becomes more severe when ownstream proucts become better substitutes (i.e. when " is higher), at the same time also the positive impact of both KT 2 an F L 1 are relatively more pronounce; the resulting ownstream price ecreases that follow from them have a larger impact on rm pro ts when " is higher. These two positive e ects consistently outweigh the negative e ect of P E 4 when " > 2. Finally, in orer to consier the intra-inustry e ects of a change in, we have to consier both = with MNEs ownstream, as well as u = with MNEs upstream, for in practice MNEs will simultaneously serve as ownstream (customer) rms for some local companies, an as upstream (supplier) rms for others. First consier = with MNEs ownstream: We see that in both the extreme cases ( = 0; 1) its sign is unambiguously positive. The reason is that in both cases, one of either two positive e ects in (12) is maximize, which more then compensates the remaining negative e ect of P E 2. Speci cally, when = 0, KS 1 is maximize an when = 1, IDL 1 is maximize. For u = with MNEs upstream, we see a conitional positive e ect when = 0 an an unconitional negative e ect when = 1. The latter is obvious: If = 1, KS 2 in (13) is zero, so that only the negative e ects of IDL 2 an P E 3 remain. When = 0, the negative e ect of IDL 2 isappears an the e ect of KS 2 is maximize. The conition states that the larger xe costs are relative to marginal costs (or more precisely: the more important the e ect of knowlege i usion on xe rather than marginal costs) the more likely it is that u = > 0. The reason is that the negative price e ect P E 1 only works through marginal costs, whereas the positive knowlege spillover e ect KS 2 works both through xe an marginal costs. Taken together, these e ects imply that horizontal (intrainustry) e ects of increasing are positive if = 0 an xe costs are large relative to marginal costs, whereas they are ambiguous if = 1. In the remainer of this paper we will empirically explore these insights regaring the horizontal an vertical e ects of MNE ownership on local rms. Moreover, the cross-country nature of our rm panel also allows us the empirically investigate the erive e ects of i erences in IPP. 4 Data an methoology 4.1 Methoology Although our theoretical moel erives preictions regaring the relationship between rm pro ts an MNE ownership, the extant literature on knowlege i usion from FDI usually consiers the e ect of MNE presence on local rms prouctivity. In orer to enhance comparability of our results, we also follow this approach in the empirical section of the paper. Moreover, from the pro t functions in Section 2 it is clear that there exists a positive an proportional relationship between rm prouctivity an rm pro ts. 13

15 The empirical moel that we will estimate takes the following generic form:! ijkt = Horizontal jkt + 2 Backwar jkt + 3 F orwar jkt (15) + 4 X it + D j + D k + " ijkt where i, j, k an t inex rm, inustry, country an time (year) respectively,! is rm level prouctivity, Horizontal is a measure of intra-inustry MNE presence, Backwar (F orwar) is a measure of MNE presence in customer or ownstream (supplier or upstream) inustries, X is a vector of rm level control variables, D j an D k are two sets of inustry an country ummies, " is an error term which is clustere at the inustry level an assume to be normally istribute, an the s are the parameters to be estimate. The precise measurement of these variables is explaine below. A well-known problem with empirical moels such as the one in (15) is the measurement of the epenent variable. Prouctivity is usually compute as the error term of a prouction function. However, to the extent that (expecte) changes in prouctivity are observe or anticipate by rms managers, the requirement of inepenence between the error term an the inepenent variables is violate, since managers may ajust variable inputs an prouction factors (such as labor) in anticipation of prouctivity changes. Olley an Pakes (1996) suggest a robust estimator to tackle this issue. The unerlying iea is that there exists a relationship between unobserve prouctivity on the one han, an observable investment an capital on the other han. Using this relationship, one can control for prouctivity in the prouction function estimation, by aing the function of investment an capital in aition to labor an capital (an material) inputs. 7 Levinsohn an Petrin (2003) exten this approach to situations in which there are a lot of zero observations on rm level investment, in which case it is not possible to invert the investment function, an hence to erive the prouctivity function. Since virtually all of our rms have positive observations on investment, we will use the Olley an Pakes (1996) proceure to estimate prouctivity. 8 One important aitional issue we nee to tackle is the fact that our theoretical moel oes not suggest a clear functional form regaring the relationship between MNE ownership in its subsiiary an local rm pro ts (an prouctivity). This can be note from the expressions of the total erivatives (cf. Appenix), which themselves are polynomial functions of MNE ownership. Moreover, the egree of the polynomial in epens on the elasticities of substitution an eman. A secon issue in this regar is that some parameters in our moel may be functions of themselves, such as the i erent elasticities of substitution or the input an output shares, which further inuces the i erent total erivatives to be (polynomial) functions of. Hence it woul be inappropriate to specify a functional form empirically ex ante. Fortunately, we can use 7 Since the appropriate functional form of the function of investment an capital is not known, Olley an Pakes (1996) use a thir-orer polynomial expansion in both variables to proxy the function. We follow this proceure in our prouction function estimation. 8 In orer to empirically implement the estimator, we use a programme evelope by Arnol (2003). 14

16 (semiparametric) partial linear regression analysis to get a clue regaring the proper empirical speci cation. Speci cally, the generic partial linear regression moel in our case takes the following form: y i = g(z j ) + X i + ij (16) where i an j again inex rm an inustry respectively, g(z) is the nonparametric component of the moel for which the functional form is etermine using a Kernel estimator, X is a vector of ( rm level) variables that enter the moel in the usual parametric fashion, an is an error term. In the context of the present paper, the variables measuring MNE presence woul enter the non-parametric component, whereas the control variables enter the parametric component. The moel in (16) can be estimate by using a i erence estimator (Robinson, 1988). Lokshin (2006) proposes the following estimator of the moel, base on Yatchew (1997), where the observations are orere as z 1 < z 2 < :: < z N : mx mx k y i k = k g(z i k 1 k 1 k ) +! mx mx k x i k + k i k (17) k 1 k 1 where m is the orer of i erencing an the s are the i erencing weights. 9 When optimal weights are chosen, OLS estimation can be consistently applie to (17) in orer to obtain estimates for the parametric part of the moel. If we enote the resulting estimator by ^ iff we can retrieve the nonparametric component in (16) as follows: y i ^ iff x i = g (z j ) + ( ^ iff )X i + " i ' g(z j ) + i We can then use a nonparametric estimator to estimate the nonparametric component g(z j ). Here we again follow Lokshin (2006) who proposes the use of a Locally Weighte Scatterplot Smoother (lowess). Lowess belongs to the class of Nearest Neighbours Estimators: It estimates local polynomials to erive a functional form for g(:), base on the istribution of the observations in a zyscatterplot. The local polynomial estimation is repeate over small parts of the istribution, where the partitioning (in so-calle banwiths) is variable. This results in a smoothe t of the relationship between z an y, which can be epicte in zy space. Finally, we nee a way to etermine whether or not the nonparametric component in (16) makes a signi cant contribution to the moel. Obviously, since we are not estimating any parameter values, we cannot use regular test statistics to etermine signi cance. Instea, Lokshin (2006) propose the following test statistic: V = p mn(s 2 res s 2 iff )=s 2 iff N(0; 1) (18) 9 These weights have to satisfy two conitions: (i) P m k 1 k = 0, which assusres that the nonparametric component in (16) is remove since g() is assume to be smooth, single-value an to have a boune rst erivative; (ii) P m j 1 2 j = 1 which assures that the resiuals in (16) have variance 2. 15

17 where s 2 res is the mean square resiual of the parametric regression an s 2 iff the square resiual of the semi-parametric regression. Hence, if this test statistic surpasses the stanar normal critical values at usual signi cance levels, we can conclue that the nonparametric component contributes signi cantly to the moel in (16). Despite the attractive property of not having to specify an explicit functional form between prouctivity e ects an MNE ownership, there are some other caveats of partial linear regression analysis. The most important of these is that the metho of Lokshin (2006) is only applicable to cross-section samples, so that we loose a lot of information containe in the time-series imension of the ata. The secon rawback is that this metho oes not allow for clustering of the error term, which is problematic when estimating rm-level prouctivity e ects while using rm an sector level explanatory variables. Thir, because of the nee for a fairly large sample to consistently estimate the partial linear moel (the so-calle curse of imensionality in semiparametric an non-parametric regression analysis), it is unwarrante to split up the sample accoring to IPP levels, as this woul heavily reuce the size of the resulting subsamples. Because of these rawbacks, we use the semiparametric approach mainly for exploratory purposes, an revert to a more stanar parametric speci cation to tackle these three issues. Summarizing, in orer to obtain a proper estimate of our epenent variable in moel (15) we use the Olley an Pakes (1996) proceure. Moreover, since we have no clear theoretical inications regaring the proper functional form of the relationship between rm prouctivity an MNE ownership, we use semiparametric regression analysis to n the best parametric speci cation for this relationship. We will then take the functional forms suggeste by the partial linear regression moels an impose it in a stanar parametric regression moel like the one in (15). 4.2 Data Our sample contains a short panel of 1549 large, publicly trae rms that are active in 20 countries an 18 sectors uring the perio Of these rms, 327 are partly owne by an foreign company. In orer to obtain the prouction function parameters with the Olley an Pakes (1996) proceure, we estimate prouction functions at the two-igit ISIC Rev. 3 level. A full list of countries an sectors is inclue in the Appenix. Our main variable of interest, i.e. the extent of intra-inustry MNE presence, is compute as follows (cf. Javorcik, 2004): Horizontal jt = P nj i=1 ( i Sales it ) P Nj i=1 Sales it s.t. 0 i < 1 (19) where n j is the number of foreign-owne subsiiaries present in sector j, N j is the total number of rms in sector j, i is the share of MNE ownership in the subsiiaries, an Sales i are the amount of rm-level sales. As with most 16

18 Backwar jt = X k6=j( jk Horizontal kt ) (20) where Horizontal k is e ne as in (19). Hence, in line with the theoretical moel evelope in Section 2, the extent of backwar linkages is proxie by the amount of inter-inustry sales from inustry j to k. Forwar linkages are compute in an analogous manner: F orwar jt = X j6=k( jk Horizontal kt ) (21) empirical stuies using MNE ownership, we only have observations for in one year (2004), which we also use to compute Horizontal in the other years. In line with Javorcik (2004), we use input an output shares (constructe from OECD I-O tables) to compute forwar an backwar linkages. 10 Speci - cally, if jk enotes the output share of sector j owing to sector k (with j 6= k) backwar linkages are compute as: where jk is the share of inputs that sector j obtains from sector k. Javorcik (2004) nets out exports from the host country to other countries from Hoirzontal kt in this case,.since such exports are obviously not estine for local sector j. However, ue to lack of ata we are not able to follow this approach, an have to settle with the computation in (21). As explaine in the previous section, our epenent variable is the prouctivity of local rms, compute using the Olley an Pakes (1996) methoology, an using ata on net sales an revenue, employment, net xe capital stocks an total investment for the years We a two control variables: First, we use a measure of rm size, measure by (the log of) total assets of the rm. The expecte sign of this variable is unclear: Some authors have argue that large rms are conucive to innovation an hence prouctivity, because of economies of scale (Cohen an Klepper, 1996). Yet others argue that resources are not easily an e ciently allocate in large rms, hence wasting prouctive resources an ecreasing prouctivity (Acs an Auretsch, 1990). The sign of this variable is thus an empirical matter. Secon, in orer to also incorporate a relative measure of rm size, we use the share of rm-sales in total inustry-sales (i.e. market share) as an aitional control variable. Again, the sign of this variable is not clear ex-ante. Table 2 below presents some summary statistics an pairwise correlations between the variables. << INSERT TABLE 2 ABOUT HERE >> We also have to construct variables that enable us to test the conitions in Table 1. For this purpose, we follow earlier research (Javorcik, 2004b; Allre an Park, 2007) an use the Ginarte an Park (1997) ataset containing ata 10 Although our ata pertain to the perio , the most recent I-O tables available are from 2002, so that we use these ata to compute input-output shares for the entire perio. 17

19 on the strength of national IPP systems. 11 The most recent set of observations relate to the year 2000, which are the ones we have use in the empirical part of the paper. The IPP inex is mae up out of ve i erent components, all rate on a 0 to 1 scale (cf. Ginarte an Park, 1996 for a etaile escription of this inex). Taken together, the IPP inex is measure along a 5 point scale, where a value of 0 inicates very weak IPP an 5 inicates very strong IPP. Regaring the horizontal prouctivity e ects, we note that they are more likely to be positive uner = 0 when xe costs make up a relative large share of total costs (i.e. xe an variable costs). We use ata on net xe assets F (property, plant an equipment) to capture rm-level xe costs, an ata on salaries an bene t expenses L to capture variable costs. We then construct a variable (F + L)=L which correspons with the conition in Table 1. For forwar prouctivity e ects ( = with MNEs upstream) we establishe that if the elasticity of substitution " between ownstream proucts is large enough, this e ect will be positive. If we interpret " as a measure of ownstream competition (with higher " inicating more substitution an hence more competition) we can construct a Her nahl inex to measure the inverse of ". Hence, for each country-inustry-year combination in our sample, we construct a Her nahl inex which captures all our sample- rms which belong to a particular sector. Finally, the conition regaring backwar prouctivity e ects ( u = with MNEs ownstream) epens inter alia on an. is alreay incorporate in the computation of (20). Since we o not have the ata to compute proper estimates of, we will just focus on in the empirical part of the paper. 5 Empirical results 5.1 Semiparametric results Before turning to the regression results, we rst brie y consier the prouctivity i erence between local rms an s, since the presume prouctivity superiority of MNEs an hence s vis-à-vis local rms lies at the heart of our moel, an as such at the hart of the knowlege i usion process. Comparing the log of prouctivity levels of the 327 s in our sample versus the 1222 local rms, the former have an average prouctivity of 5.80 an the latter A paire t-test strongly rejects the equality of these two means (t = 14:3). Hence, the superiority of s with respect to local rms on prouctivity as assume in our theoretical moel is con rme in our sample. First we consier the results of the semiparametric partial linear regression moel. We will investigate the e ect of the three i erent MNE presence variables separately, in orer to obtain the empirical functional relationship between prouctivity an the relevant MNE ownership share. As explaine in the previous section, the partial linear regression estimator we use is only applicable in cross-sections. Thus all results reporte in this subsection pertain to the year 11 We thank professor Park for sharing the upate ataset. 18

20 2004, which is the year in which the MNE ownership shares were observe. The results of the partial linear regression moel are reporte in Table 3 below. Figure 2 contains the resulting non-parametric relationship between prouctivity an each of the MNE presence variables. The rst column in Table 3 as the horizontal variable from (19) to the non-parametric component of the moel. As inicate by the test statistic V from (18), the non-parametric component enters the moel highly signi cantly. Panel (a) in Figure 2 epicts the implie relationship. We n that an increase in MNE ownership in the increases local rms prouctivity. However, it is also clear that this relationship is not linear, but characterize by ecreasing returns to MNE ownership at low levels of MNE ownership, an increasing returns to MNE ownership at high levels. Hence, the semiparametric moel suggests a cubic relationship between intra-inustry MNE ownership an local rms prouctivity. << INSERT TABLE 3 ABOUT HERE >> << INSERT FIGURE 2 ABOUT HERE >> In the secon column we put the backwar variable from (20) in the nonparametric component. The test statistic V again inicates that the nonparametric component enters the moel highly signi cantly, an panel (b) in Figure 2 epicts the relationship between the ownstream MNE ownership share an upstream local rms prouctivity. The gure emonstrates a quaratic relationship, although the 95% con ence interval aroun this relationship is quite large. Finally, in column three of Table 3 we put the forwar variable from (21) in the non-parametric component of the moel. Forwar spillovers enter the moel highly signi cantly an from panel (c) in Figure 2 we see that the relationship between upstream MNE ownership an ownstream prouctivity of local customers is again characterize by a quaratic relationship. But also in this case, the 95% con ence interval is rather wie. Both rm size an market share are signi cant an positive, inicating that both absolute rm size as well as rm size relative to the market are conucive to prouctivity. In terms of moel t, the moels perform rather well, inicating that the inustry an country xe e ects also absorb a lot of the variation in rm prouctivity. However, in orer to tackle the three problems escribe in the previous section, we have to revert to parametric regression analysis. In oing so, we can use the outcomes of the semiparametric moels as guie regaring the parametric moel speci cation. Speci cally, the semiparametric results suggest that we nee quaratic an cubic speci cations to capture the relationship between rm prouctivity an MNE ownership. Hence, we construct two new variables: P nj Horizontaljt 2 i=1 = (2 i Sales it) P Nj s.t. 0 i < 1 (22) Horizontal 3 jt = i=1 Sales it P nj i=1 (3 i Sales it) P Nj i=1 Sales it 19 s.t. 0 i < 1 (23)