Multinational …rms, Technology and Location PehrJohan Norbäck The Research Institute of Industrial Economics¤ June 19, 2000
Abstract This paper analyzes a threestage optimization problem in which a …rm chooses (i) its technology, by deciding on a level of R&D, (ii) whether this technology is to be used in a domestic or a in foreign plant and (iii) the quantity produced and sold on the market. If technology transfer costs are low, “hightech” or R&Dintensive …rms tend to produce abroad. At higher technology transfer costs, hightech …rms tend to export. An empirical analysis using a data set of Swedish multinational …rms, con…rms the latter prediction. Keywords : Multinational Firms, R&D, Location, Empirical Analyses JEL classi…cation: L13, F23, O33 ¤
PehrJohan Norbäck, the Research Institute of Industrial Economics Box 5501 SE114
85 Stockholm. Email:
[email protected], Internet: http://www.iui.se, Telephone: +46 8 665 45 22 and Telefax: +46 8 665 45 99.
1
1
Introduction and summary
In the literature, multinational …rms (MNFs), that is, …rms performing economic activities in multiple countries, are closely related to …rmspeci…c assets.1 Firmspeci…c assets, which include such things as marketing ability, product di¤erentiation or Research and Development (R&D), can be seen as giving a …rm a competitive edge, which enables it to expand production into foreign markets. Recent imperfect competition models of multinationals also show that …rms are more likely to choose foreign direct investment (FDI) when …rmlevel …xed costs, such as R&D expenditures, are high, relative to plant level …xed costs. Seminal papers include Horstmann and Markusen (1992), Brainard (1993), Ethier and Markusen (1996) and Markusen and Venables (1998). These models typically treat …rmspeci…c …xed assets as …xed visavis the entry choice into a foreign market. However, when analyzing the relationship between R&D and FDI, say, this assumption overlooks the fact that a …rm may not only expand sales abroad in order to draw on its technology asset, it may also be done to gain resources to develop this asset.2 In this paper, I therefore extend the earlier work by modelling this interaction. Quite surprisingly, this slight modi…cation of the standard model introduces an ambiguity in the relationship between R&D and FDI. Even more surprisingly, I …nd empirical evidence suggesting a negative relationship, which is quite the opposite of the traditional view. I use a framework developed from Leahy and Neary (1996). I study a monopoly …rm which makes three distinct choices: It invests in costly R&D to improve its technology, thereby decreasing the marginal cost. Then, it either implements the technology in an a¢liate which supplies the market 2
from a foreign plant, or in a domestic plant which supplies the market by export production. Given this location choice, the good is supplied. The …rm takes the fact that export production is subject to a trade cost into account. Moreover, I also assume that implementing the technology abroad is more costly, due to technology transfer costs.3 The model predicts that when transfer costs are small, hightech …rms will choose foreign direct investment. At higher transfer costs, on the other hand, hightech …rms choose to export. Hightech or R&Dintensive …rms then gain more by avoiding transfer costs of technology than by avoiding transport costs of physical units, since more complex technology demands larger resources for technology transfer. These predictions are tested on a data set consisting of Swedish multinational …rms, provided by the Research Institute of Industrial Economics (IUI) in Sweden. Both countries with foreign production and countries exclusively supplied by exports, are included in a twostage estimation procedure, where I use the share of foreign sales accounted for by overseas a¢liate production as the dependent variable in the OLSregressions. The empirical analysis shows that exports is the preferred choice by R&Dintensive …rms: There is a persistent negative correlation between R&D intensity and the a¢liate share of foreign sales, on the one hand, and the probability that any a¢liate sales are recorded, on the other. These …ndings may also be contrasted to some recent work. Contemporaneously to this paper, SannaRandaccio and Petit (1998) have developed a similar framework in which investment in R&D leads …rms towards foreign expansion but also that MNFs tend to invest more in R&D. SannaRandaccio and Petit also discuss transfer costs. On the basis that FDI predominantly occurs between developed countries, they argue that transfer 3
costs should be small and derive the twoway relationship on this presumption, predicting that hightech …rms should be predominately multinational. In contrast, this paper shows evidence of the opposite relationship: hightech …rms are more inclined to export. While the IUI data set has the advantage containing …rmlevel information, it is limited to …rms with producing a¢liates. How would then the inclusion of purely exporting …rms a¤ect the results? This may reverse the negative relationship between R&D and the probability of …nding a¢liates. However, the negative relationship between R&D and the a¢liate share will not be a¤ected since this regression, by de…nition, only includes …rms with producing a¢liates. Furthermore, it is informative make a comparison with Brainard (1997). She employs the same twostage method to investigate the pattern of US foreign production and exports, using a cross section of industries and countries. Combining trade data from the US Bureau of Census and FDI data from the Bureau of Economic Analysis, she …nds that R&D increases the probability of …nding a¢liate sales in a country. Based on her theory, R&D is, however, not included in her secondstage regression. She does however …nd that when levels of a¢liate production and exports are separately regressed against R&D intensity; both increase in R&D, but the elasticity of exports is about two and half times larger.4 She concludes that R&D is consistent with both exports and foreign production but makes no re‡ection to why the export bias is so strong. The mechanisms and results of this paper may indeed be used to explain this pattern. Some restrictions of the theoretical model should also be noted. The model uses a monopoly setup. This is purely for expositional reasons as the results in this paper also extend into oligopoly. I model R&D as costreducing, but the analysis can also be extended to quality improvements or 4
the generation of new products. I abstract from any home market in‡uence on the choice between FDI and exports. This assumption simpli…es the analysis, but does not seem too restrictive when the focus is on a country like Sweden, where the homemarket may be of neglible size for its large international …rms. For functional forms, I use linear and quadratic functions. The paper proceeds as follows: In section 2, a theoretical framework is derived. In Section 3, an empirical analysis based on the …ndings in section 2 is performed. Section 4 concludes.
2
Theory
In this section, I study the interaction between the R&D decision and the choice between foreign and export production, as alternative means of serving a market abroad.
2.1
Structure
The structure of the problem is the following. There is a single …rm producing a homogeneous good. The demand is located in another country, which may be considered as the world market for the good in question. The …rm makes three decisions: First, it invests in costly R&D at home. We assume the technological level of the …rm to be represented by its cost level of production, and that R&D lowers the marginal cost. Then, the technology is implemented either in export production from a domestic factory (henceforth denoted E), or a direct investment is made (henceforth denoted FDI) and production takes place in foreign a¢liate (henceforth denoted F). Finally, the market is supplied. For notation subscripts will denote location. For example, qh is the 5
output choice of the …rm in location h, for h = fE; Fg : Hence, qE is the
export quantity of the …rm, whereas qF is the production quantity when an a¢liate is established.
2.2
The model
The marginal cost in production in location h for h = fE; Fg, is given in (1): ch = Ch ¡ µxh;
CE = c0 + t;
CF = c0
(1)
where µ and c0 are positive constants. Several factors a¤ect the production costs. From (1), we can see that the …rm chooses levels of R&D, indicated by xh, which lower its marginal costs. Export production is also subject to a transport cost or a tari¤ barrier, t, which can be avoided by FDI. The inverse demand is given by (2): Ph = a ¡
qh s
(2)
where a > 0 is a demand parameter and s > 0 can be interpreted as a measure of the size of the market. The total pro…t can then be written as (3):
2
¦h = (Ph ¡ ch) qh ¡ °((1+±2h )xh) ¡ Gh = (Ph ¡ ch) qh ¡
(3)
° (xh)2 ¡ Th ¡ Gh 2
In (3), the …rst term indicates variable pro…ts and the last three terms indicate di¤erent types of …xed costs. From the left to the right, these …xed costs are as follows: 6
First, R&D is assumed to incur quadratic costs, so that xh gives rise to …rmspeci…c …xed costs of
°(x h) 2 2 ,
where ° is a positive constant.5 Note that
this term corresponds to …rmspeci…c costs discussed in the introduction, which are usually modelled as exogenous in the literature. Here, these costs are endogenous. Second, the …rm is assumed to have production units at home, but initiating production abroad involves additional plantlevel investments. Plantlevel …xed costs are then de…ned in (4): 8 < G; Gh = : 0;
for h = F for h = E
(4)
Third, following Teece (1977), technology transfer costs for implementing the technology in a factory located abroad are assumed to be higher. To simplify, let us normalize such that new technology can be implemented at home without cost, whereas an additional cost T of transferring the technology abroad arises, since it must be adapted to local conditions. More complex technologies may require closer interchange of information with manufacturing, thereby increasing communication and information costs if production is located abroad. The transfer cost is therefore made dependent on the actual level of R&D. This is done by introducing a parameter ±, such that 0 · ±F < 1, if foreign production is chosen and ±E = 0, if export production is chosen. It simply means that a given level
of R&D, x, equally lowers the cost of production, irrespective of location (cf. equation (1)), but that the implementation of the technology abroad requires additional R&D e¤orts of ±x. From (3), we can then restate the
7
resulting transfer costs as: 8 < ±(2+±) (x )2 ¸ 0; h 2 Th = : 0;
for h = F for h = E
(5)
where we note that T (¢) is indeed increasing with the level of R&D, x6. Pro…t maximizing production quantities, qh, and R&D expenditures, xh, are chosen so that (6) must hold: @¦h qh = Ph ¡ ch ¡ = 0; @qh s
@¦h °xh = µqh ¡ =0 @xh (1 + ±h)2
(6)
Using (2), and (6), optimal production quantities and optimal R&D levels are given by (7): qh = s
a ¡ ch ; 2
xh =
µ qh ° (1 + ±h)2
(7)
As shown by Leahy and Neary (1996), all endogenous variables can be solved in a parameter ´; de…ned as: ´´
µ2s ¸0 °
(8)
´ may then be interpreted as the relative return to R&D. Note that ´ is zero, if R&D is completely ine¤ective (µ = 0), inexcessively expensive (° = 1) or if the size of the market is very small (s = 0) :
To ensure wellbehaved solutions, we will make two assumptions: (i)
The parameter values are such that the …rm always have a strict positive marginal cost which, by (1), implies that ch > µxh holds. (ii) The parameter values support positive pro…ts in both locations. This, in turn, implies cases where ®´ < 2; where the parameter ® is a measure of the impact of the transfer cost7: 0 <®´
1 ·1 (1 + ±)2 8
(9)
By using (1), (7), (9) , I can express the optimal production level for each location choice h in ´. These are given in (10):
qE = s
A ¡t ; 2 ¡´
qF = s
A 2 ¡ ®´
(10)
where it will be assumed that A ¡ t > 0 and A = a ¡ c0 > 0.8
We can then use (3), (4), (5), (6), (7) and (10), for deriving expressions
for the total pro…ts in the two alternative locations. Then, we have: ¦E (´) = ¼E (´) ¡ TE ¡ GE 1 = (qE )2 (2 ¡ ´) 2s ¦F (´) = ¼F (´) ¡ TF ¡ GF 1 = (qF )2 (2 ¡ ®´) ¡ T ¡ G 2s
(11)
(12)
where variable pro…ts are denoted ¼h (´) and production quantities are given by (10). 2.2.1
The equilibrium location
Let us now characterize the equilibrium location of production. It is then useful to explore the variable pro…t function. It is easy to state and prove the following proposition: Proposition 1 The variable pro…t for exporting ¼ (´) is increasing and the corresponding pro…t in foreign production ¼F (´) increasing or nondecreasing in the relative return to R&D, ´: Furthermore, de…ne ® = ® ~ as a critical value of technology transfer costs. Then, a¢liate pro…ts ¼E (´) increase at a faster rate in ´, compared to export pro…ts ¼E (´) if transfer cost are su¢ciently low, ® > ® ~ . The opposite holds if transfer costs are su¢ciently high, ® < ® ~. 9
Proof. See appendix A What is the economic intuition behind proposition 1? First, the pro…t in export and foreign production increases in ´ simply because a higher return to R&D implies higher spending on R&D, thereby lowering the marginal costs. Moreover, since FDI avoids the transport cost, larger sales in foreign production also imply increased spending on R&D, as compared to the alternative of exports. Therefore, the di¤erence in marginal costs between the two location alternatives will tend to exceed the transport cost t, and this di¤erence will be increasing in ´. This tradecoste¤ect will work towards making pro…ts in a¢liate production more responsive to a increase in the relative return to R&D ´. However, locating production abroad implies that technology need to be transferred from home R&D labs to abroad. At an increasing relative return to R&D ´, increasing technology transfer costs tend to restrict the …rm’s R&D which, in turn, limits the reduction of the marginal cost if a¢liate production is chosen. This transfer coste¤ect tends to make pro…ts in export production more responsive to a increase in the relative return to R&D ´ Which location of production is then actually chosen? Comparing total pro…ts in the two alternatives, the following proposition applies. Proposition 2 Suppose that parameter values are such are such that total pro…ts are equal in export and a¢liate production at some ´¤ . If technology transfer costs are low ® > ® ~ , then a …rm endowed with a high relative return to R&D, ´ > ´ ¤ chooses FDI, whereas the …rm exports if ´ < ´ ¤: The opposite holds if technology transfer cost are high, ® < ® ~. Proof. See appendix A 10
Accordingly, the model predicts that when transfers of technology is less costly, …rms in knowledgeintensive industries, that is, industries with a relatively high relative return to R&D ´, are inclined to locate production abroad, while …rms in industries with a lower return to R&D choose to export. On the other hand, when it is more costly to transfer technology, hightech …rms tend to export whereas lowtech …rms choose FDI. In the latter case, hightech …rms gain more by avoiding transfer costs than by avoiding transport costs of physical units. These results are summarized in table 1. The table also shows comparative statics result for both cases of transfer costs. The explicit expressions are given in appendix A. The …rst column indicates an increase in the exogenous variable z. The second and third columns reveals the e¤ect in the case of low transfer costs. More speci…cally, the second column shows the qualitative e¤ect on ´ ¤; whereas the third column translates this into the “marginal e¤ect” on the …rm’s incentive to choose FDI and locate production abroad9 . This sign can be interpreted as the e¤ect on the location decision in a marginal …rm endowed with a relative return to R&D of ´ ¤. Column three and four does the same thing for the case with high transfer costs. The comparative statics results in table 1 reveal no surprises, so I will be very brief in commenting on them. Whatever the size of transfer costs, FDI is less likely when plantlevel …xed costs G are higher and more likely when transfer costs are lower (® is larger) and when trade barriers t are higher. For example, in the low transfer cost case, an increase in t will lower ´ ¤ and induce the marginal …rm to produce abroad. In this case, the marginal cost in a¢liate production does not only decrease due to increased transport costs, but also due to the fact that a more extensive production increases 11
R&D expenditures. These e¤ects magnify the di¤erence in marginal costs between export and a¢liate production, which, in turn, allows pro…ts in these two alternatives to be equalized at a lower return to R&D ´ ¤. Hence, higher transport costs favor FDI, since a larger range of ´ permits direct investment.
3
Empirical analysis
The theoretical section gives an ambiguous view of the relation between a …rm’s technology and its choice between a¢liate and export production. Since di¤erent predictions arise depending on the importance or level of technology transfer costs, this provides an opportunity to test the impact of technology transfer cost.
3.1
Data
The primary data source is a data set from the Research Institute of Industrial Economics (IUI), based on a questionnaire sent to all Swedish MNFs every fourth year, on average. Data is available from seven surveys: 1965, 1970, 1974, 1978, 1986, 1990 and 1994. The survey covers almost all Swedish multinational …rms in the manufacturing sector, and detailed information is available on variables such as R&D, employment, production and their distribution between domestic and foreign units, as well as on internal and external trade ‡ows. This rich data set has been used in the following way: (i) All …rms with at least one production a¢liate abroad are included in the sample. (ii) Within this set of …rms with production a¢liates, we focus on foreign sales to the OECD countries.10 (iii) All exports sales are sales of …nal goods, that 12
is, the impact of input goods is removed. (iv) Exports back to Sweden from the a¢liates have been removed from a¢liate production. Let me brie‡y comment on these conditions. Ideally, …rms without production a¢liates should be included in the sample, but corresponding …rmlevel data for purely exporting …rms is simply not available. I have chosen to focus on OECD countries, since the modelling framework does not emphasize di¤erences in factor costs. In addition, sales to OECD countries cover the vast majority of foreign sales in these …rms. Finally, the last two criteria are chosen to comply with the absence of inputgoods and homemarket e¤ects in the theoretical section. Additional information on country and industry speci…c variables are taken from World Development Indicators (1997), OECD (1997) and SCB.
3.2
The econometric model
The theory presented in the previous section predicts a …rm’s choice between implementing its technology in export or a¢liate production. In translating this theoretical prediction into an empirical analysis, there are several caveats: The simple model involves a …rm which produces a single …nal product, whereas many of the …rms in the sample are large multia¢liate …rms with multiple product lines. Furthermore, within such …rms, the location of technology cannot be directly observed from the data. I will follow Brainard (1997) and use the share of foreign sales accounted for by the a¢liates as my dependent variable. This variable is labeled AF SH AREijt , and is de…ned: AF SH AREij t =
SQijt SQij t + SXijt
(13)
where SQijt denotes the level of production for …rm i in country j at time t and SXijt is the corresponding export level. This relative measure then 13
indirectly captures the implementation choices of the …rms since the sales pattern must re‡ect location choices. It also subsumes the two endogenous variables, export and a¢liate production, into a single variable. The dependent variable in (13) is censured  it can take on any values between zero and one. A closer look at the data set reveals that the …rms only have production a¢liates in a minority of the countries for which foreign sales are recorded. Thus, AF SHAREijt contains a large number of zeros. Omitting these observations will result in a systematic selection bias causing any OLSestimates on AFSHAREijt to be both biased and inconsistent. Therefore, I will use a twostage procedure. This procedure, given by (14) and (15), separates the probability and marginal e¤ects: DAF SH AREij t = ¯ 0 + ¯ 1 RDit + ¯ 2T REMBht + ¯ 3 DISTj (14) +¯ 4 AGE1it + ¯ 5 RD2 it + ¯ 6GSCALE1ht +¯ 7 GDPjt + Àijt
AFSHAREijt = ®0 + ®1RDit + ®2T REMBht + ®3DISTj
(15)
+®4 AGE2ijt + ®5RD2it + ®6GSCALE2ht +®7 GDPjt + ®8¸ijt + "ijt I …rst estimate the probability of …nding a¢liate sales in a country, using the dependent variable DAFSHAREijt , where DAF SH AREijt = 1 if AFSHAREijt > 0, DAF SH AREijt = 0 otherwise. Then, a twostage selection biased corrected regression model from Heckman (1979) is employed, where the error correction variable ¸ijt is included. The explanatory variables are presented below. Logs are also used in all continuous variables.
14
3.3
Exogenous variables
In table 1, the independent variables and the corresponding exogenous variables from the theoretical section (for which they act as proxies), are presented. For convenience, I also reproduce their expected sign, based on my …ndings in the theoretical section. Two kinds of independent variables will be used; core variables and additional variables. 3.3.1
Core variables
This group of independent variables is closely attached to the exogenous variables encountered in the theoretical section. R&D intensity, RDit , de…ned as the share of R&D expenditures in the total sales of the …rm, is used as a proxy for the relative return to R&D, ´. Since the focus in this paper is on the relation between technology and location, RDit is the explanatory variable of most interest.11 . Note that since R&D expenditures are endogenously determined in the theoretical section, there may be an endogenity problem with the R&D variable in the empirical analysis. The structure in our theoretical models  in which R&D expenditures are set before location decision and market interaction  suggests that a lagged R&D intensity should be considered. Ravenscraft and Scherer (1982) propose a lag of approximately …ve years between R&D expenditures and pro…ts, which suggests that the fouryear lag, corresponding the one period lag in terms of surveys, should be treated as endogenous. To reduce any simultaneity bias, the eight year lag on R&D intensity will be used instead12. However, since two surveys are lost in the lag procedure, and because of unbalanced nature of the data set as many …rms disappear when they are acquired or reorganized over time, I will also report estimations, using the
15
present R&D intensity. This avoids a massive loss of observations associated with the eightyear lag. Note also that given that R&D is conducted before any market interaction, R&D in time t should be uncorrelated with the error terms in (14) and (15). Unfortunately, no direct measure of the plant level …xed costs G can be calculated, as the data base lacks information on individual plants in the Swedish part of the corporation. Information on plant size is available for a¢liates, but using this information without care gives rise to two immediate problems: (i) If plantlevel scale economies are su¢ciently large, then we would suspect that domestic production is preferred, thereby indicating that proxies for G based solely on a¢liate information may be misleading. (ii) Relating large a¢liate plants directly to AF SHARE may give a spurious correlation  large a¢liates should account for a large share of foreign sales, a relationship which may have little to do with the e¤ect of scale economies on the location decision. In the probit equation (14), I will use GSCALE1it . It is de…ned as the ratio between the average number of employees in a¢liates and the average number of employees in the corporations. To reduce the above problems, I aggregate to the three or fourdigit industry level. The OLS stage should be more sensitive, however, since it directly uses the continuous variable AFSHARE; whereas the probit stage involves the dichotomous variable DAFSHARE. Therefore, equation (15) uses GSCALE2it , which is instead calculated from Swedish industrial statistics. It is de…ned as the average size of plants with more than one hundred employees divided by total industry mean size, at the three or fourdigit industry level to which the …rm belongs. Turning to measures of transport costs, T REMBit is calculated as the share of transport and packing costs in total variable costs, and once more, 16
Swedish three or fourdigit industry level data are used. In addition to packing and transport costs, total variable costs include costs for electricity, raw materials and wages for bluecollar workers. The second measure, DIST Wjt , is an index measuring the geographical distance from Sweden to the respective countries. It is very di¢cult to …nd a variable which accurately captures the effect of technology transfer costs. Following Swedenborg (1982), it may be argued that more experience of production abroad should lower technology transfer costs to units abroad, and that this should also be the case for …rms performing R&D abroad13. To capture the e¤ects of experience in foreign production, AGE1it re‡ects a weighted average of the age of the a¢liates of a …rm, irrespective of their location. AGE2ijt is simply the mean age of the a¢liates in a particular country. To measure the e¤ects of R&D abroad, we construct two dummy variables; RD1it takes on the value of one, if the …rm performs any R&D abroad, and RD2ijt , which takes on the value of one if the …rm performs any R&D in the country in question. 3.3.2
Control variables
In addition to the core variables, a set of control variables will also be included. The …rst control variable is the size of the respective country measured as PPPadjusted, de‡ated GDP, GDPjt .14 Following Brainard (1997), I also control for di¤erences in factor proportions through the variable IN COMEjt , measuring per capita income di¤erences between Sweden and the respective countries where the …rm operates. OPENjt is an openness index taken from Wheeler and Moody (1992), measuring the openness of a country to FDI. Finally, following Braunerhjelm and Svensson (1996),
17
I control for the in‡uence of agglomeration e¤ects on the location decision through the variable AGGLOMhjt . This variable is de…ned as the share of all employees in the manufacturing industry in the industry to which the investing …rm belongs, out of all employees in the manufacturing sector in the respective countries, divided by the share of employment in this sector in all countries. If pecuniary externalities in terms of cost and demand linkages are present in an industry, thereby attracting direct investments, such agglomeration forces should be captured by this variable.15 Finally, I also control for speci…c e¤ects over regions, industries and time. The regions are EFTA, the EC, North America and the Far East. Industry dummies are employed at the two or threedigit industry level.
3.4
Estimation results
Table 3 reports the results from estimating (14) and (15). Speci…cation (i), uses the contemporaneous R&D intensity, whereas speci…cation (ii) uses the eightyear lag on R&D intensity. The …rst two columns shows the probability and marginal e¤ects, whereas the third column shows a 2SLS estimation on equation (15) where the error correction variable is omitted. The contempemperaneous R&Dintensity is used as instrument for the four year lag in R&Dintensity in speci…cation (i). The eightyear lag on R&D intensity is used as an instrument for the four year lag in R&Dintensity in speci…cation (ii). Irrespective of speci…cation, both the probability e¤ect and the marginal e¤ect are signi…cantly negative for RDINT . This is also the case for the 2SLS estimates. That is, the larger the R&D intensity, the smaller the probability that a …rm locates production in a country and  given that
18
production is established  the smaller is the share of foreign sales accounted for by the a¢liates. Note that this negative, signi…cant sign lends supports the relationship predicted when transfer costs were assumed to be high, thereby rejecting the prediction when transfer costs were assumed to be low. Turning to transport costs, TREMB has the predicted, positive sign in all equations. DIST is statistically signi…cant, but appears with di¤erent signs in the probit and the regressions. Thus, when the geographical distance increases, the probability of a …rm locating production in a country decreases, whereas  given that a¢liates are established  a larger distance favors local production. The latter result is predicted in the theory section, whereas the former is somewhat unexpected. As Ekholm (1998) argues, it may be the case that a larger distance also re‡ects cultural and institutional factors, in which case the increasing cost of FDI dominates the e¤ects of transport costs. Variables AGE1; RD1 and GSCALE1 all have the predicted signs in the probit stage in both speci…cations. Hence, more experience in foreign production clearly increases the probability of producing abroad, whereas scale economies at the plantlevel work in the opposite way. Turning to the regressions, the corresponding variables, AGE2; RD2 and GSCALE2, reveal similar information. Thus, if the …rm has established R&D laboratories in a host country, this obviously facilitates transfers of technology and production to such a country. We also note that a sample selection bias indeed exists, as the coe¢cient on LAMBDA is positive and highly irrespective of speci…cation. The GDPvariable GDP exerts a signi…cant positive in‡uence  the size of a country is of great importance for a …rm’s decision to establish produc19
tion. Turning to the secondstage regressions, once a¢liates are established, local production seems be chosen over exports to a larger extent, when country size increases. Overall, the OLS and 2SLS estimations provide quite close results. The only exception is the GDPvariable, where the 2SLS estimations suggest that …rms on the margin favors to export to large countries rather than produce locally. This is somewhat surprising, but suggests that size of country has a stable in‡uence on a …rm’s decision to establish production, whereas the e¤ect is less clear once production has already been established.16 3.4.1
Robustness of results
Table 4 checks the robustness of the results in table 3 by …rst adding more control variables. Speci…cation (iii) reestimates (14) and (15), adding the three new variables: IN COME, OPEN and AGGLOM. The parameter estimates are quite robust to the inclusion of the extra variables. In particular, the negative relationship between R&D intensity and foreign production, as measured by DAFSHARE and AFSHARE, persists. The coe¢cients on INCOME, proxying for di¤erences in relative factor endowments, does not seem to a¤ect the probability of …nding a¢liates, but is positive and signi…cant in the OLS. Hence, within this set of OECD countries, factor proportions seem to explain some of the variation in the dependent variable once a¢liates are established: The openness of a country to FDI has the positive predicted sign, but is not signi…cant. Finally, pecuniary externalities in the shape of cost and demand linkages in the host countries seem important, as AGGLEM signi…cantly increases both the share of foreign sales of the a¢liates and the probability of establishing
20
production. Next, I undertake panel estimations controlling for …rmspeci…c e¤ects, since unobservable …rmcharacteristics may drive the results. In speci…cation (iv), I reestimate augmented versions of (14) and (15) separately, controlling for …rmspeci…c, timespeci…c and countryspeci…c e¤ects. To allow for …xed e¤ects, I use a logit formulation in (14). Since there is no withinvariation in the distance variable, DISTj and the openness measure for FDI, OP ENj , these variables are dropped. I choose to use the present R&D intensity in all speci…cations. This allows me to exploit the full timeseries variation of the data. It also avoids a large loss of observations associated with using lagged R&D as the data is heavily unbalanced. The panel results in speci…cation (iv) produce no drastic changes. In particular, there is again a negative correlation between R&D intensity and the propensity to produce abroad as measured by either the a¢liate share of foreign sales or the probability that any a¢liate sales are recorded. This is also the case for speci…cation (v), which controls for …rmcountrypair …xed e¤ects and timespeci…c e¤ects.
4
Conclusions and discussion
In the literature, high R&D intensity is often associated with multinational …rms. It should then be expected that higher R&D intensity should lead …rms to choose overseas production relative to exporting. In this paper, I have shown that this is not necessarily the case. On the contrary, using a unique data set of Swedish multinational …rms, I …nd that R&D intensity is negatively related to the share of foreign sales accounted for by the a¢liates on the one hand, and the probability of …nding a¢liate production in a country on the other. That is, on the margin, there is a negative relationship 21
between technology and overseas production. I have suggested that one way to explain this puzzle is to take into account that it is costly to transfer technology abroad. R&D intensive …rms may then …nd it more pro…table to avoid technology transfer costs rather than physical transport costs. This also points to a natural extension of the analysis, since a way for a …rm to avoid technology transfer costs associated with foreign production could be to place the development of its technology abroad. Based on earlier studies of Swedish MNFs, however, adapting technologies to local conditions and regulations rather than developing new technology, seems to have been the primary motive for locating R&D abroad.17 But it should be noted that the average share of R&D performed abroad has increased from around 9 % in the early observations to 25 % in the latest. While this shows that the major part of R&D still takes place at home, the increasing share of foreign R&D might also indicate that R&D to a larger extent takes place abroad to develop new products and technologies. Future research should therefore investigate the interaction of the production and R&D location more closely.
22
Acknowledgement I am grateful to Harry Flam at the Institutute for International Economics at Stockholm University (IIES), for useful discussions and suggestions. I am also grateful to Jim Markusen at University of Colorado at Boulder for his comments and suggestions. At the Research Institute of Industrial Economics (IUI), I want to express my gratitude to Roger Svensson, Karolina Ekholm, Pontus Braunerhjelm, Jörgen Nilson, Per Thulin, Erik Mellander, Petter Lundvik, KarlMarkus Modén, Henrik Braconier and Mattias Ganslandt. The comments made by two referees also signi…cantly improved the paper. Finally, I take the full responsibility for any remaining errors or misinterpretations in text.
23
References Blomstrom, M. and Y. Wang, 1992. Foreign investment and technology transfer: A simple model, European Economic Review 36, 137–155. Brainard, S.L., 1993. A simple theory of multinational corporations and trade with a tradeo¤ between proximity and concentration, NBER Working Paper No. 4269. Brainard, S.L., 1997. An empirical assessment of the proximityconcentration tradeo¤ between multinational sales and trade, American Economic Review 87, 520–544. Braunerhjelm, P. and R. Svensson, 1996. Host country characteristics and agglomeration in foreign direct investment, Applied Economics 28, 833–40. Caves, R.E., 1996. Multinational Enterprise and Economic Analysis. 2nd edition, (Cambridge University Press, Cambridge and New York). Cheng, L., 1984. International competition in R&D and technological leadership, Journal of International Economics 17, 15–40. Dunning, J.H., 1977. Trade, location of economic activities and the MNE: A search for an eclectic approach, in P.O. Hesselborn, B. Ohlin and P.M. Wijkman, eds., The International Allocation of Economic Activity, (MacMillan, London), 395–418. Ekholm, K., 1998. Proximity advantages, scale economies and the location of production, in P. Braunerhjelm and K. Ekholm, eds., The Geography of Multinational Firms, (Kluwer Academic Publishers, Boston), 59–76. 24
Ethier, W.J. and J.R. Markusen, 1996. Multinational …rms, technology di¤usion, and trade, Journal of International Economics 41, 1–28. Fors, G., 1997. Utilazation of R&D results in the home and foreign plants of multinationals, Journal of Industrial economics 45, 341358. Heckman, J., 1979. Sample selection bias as a speci…cation, Econometrica 47, 153–161. Horstmann, I. and J.R. Markusen, 1987. Strategic investments and the development of multinationals, International Economic Review 28, 109– 121. Horstmann, I. and J.R. Markusen., 1992. Endogenous market structures in international trade (natura facit saltum), Journal of International Economics 32, 109–129. Lall, S., 1980. Monopolistic advantages and foreign involvement by U.S. manufacturing industry, Oxford Economic Papers, 32 105–122. Leahy, D. and P. Neary, 1996. International R&D rivalry and industrial strategy without government commitment, Review of International Economics 4, 322–338. Markusen, J.R., 1984. Multinationals, multiplant economies, and the gains from trade, Journal of International Economics16, 205–26. Markusen, J.R., 1995. The boundaries of multinational enterprises and the theory of international trade, Journal of Economic Perspectives 9, 169–89.
25
Markusen, J.R. and A.J. Venables, 1998. Multinational …rms and the new trade theory, Journal of International Economics 46, 183203. OECD, 1997. Economic Outlook No 61: Annual Data (OECD Publishers). Ravenscraft, D. and F. Scherer, 1982. The lag structure of returns to research and development, Applied Economics 14, 603–620. SannaRandaccio, F. and M. Petit, 1998. Technological innovation and multinational expansion: a twoway link?, forthcoming in Journal of Economics (Zeitshrift fur Nationalokonomie). SCB, Industristatistiken. (Statistiska Central Byrån). Swedenborg, B., 1982. Swedish Industry Abroad. An Analysis of DrivingForces and E¤ects, (IUI, Stockholm). Teece, D.J., 1977. Technology transfer by multinational …rms: The resource cost of transferring technological knowhow, Economic Journal 87, 242–61. Venables, A.J., 1996. Equilibrium locations of vertically linked industries, International Economic Review 37, 341–359. Wheeler, D. and A. Mody. , 1992. International investment location decisions: The case of U.S. …rms, Journal of International Economics 33, 57–76. White, H., 1980. A heteroscedasticityconsistent covariance matrix estimator and a direct test for heteroscedasticity, Econometrica 48, 827–838. World Bank, 1997. World Development Indicators (World Bank).
26
A
Appendix:
First, the statements in propositions 1 and 2 are proved, then table 5 is derived. Finally, secondorder conditions are shown.
A.1
Proof of proposition 1
By calculation:
@¼E @´
2
= 12 s (A¡t) > 0; (2¡´) 2
@¼F @´
2
A ® = 12 s (2¡®´) 2 ¸ 0
(A.1)
which veri…es the …rst part of proposition. Then, de…ne the di¤erence in variable pro…ts as 4¼ = ¼ E ¡ ¼F . From (A.1), it is easy to see that ® = 0 (in…nite transfer costs) implies costs) implies
@4¼ @´
@4¼ @´
> 0, whereas ® = 1 (no transfer
< 0: Also note that (A.1) implies that
@4¼ @´
must be
monotonously decreasing in ®: Hence, there is a critical ® = ® ~ ; such that @4¼(´:~ ®) @´
A.2
= 0: This veri…es the second part of the proposition.
Proof of proposition 2
First, de…ne the di¤erence in total pro…ts as 4¦ = ¦E ¡ ¦F where ¦E is de…ned in (11) and (12) and note that
@4¦ @´
=
@4¼ @´ .
Then, if there is a
´ = ´¤ such that 4¦(´ ¤; z) = 0 , ¦E (´ ¤; z) = ¦F (´¤ ; z) ; where z is the
vector of exogenous variables, it follows from the proof in proposition 1 that ¦E < ¦F for ´ > ´ ¤ and ® > ® ~ , whereas ¦E > ¦F for ´ > ´ ¤ and ® < ® ~:
27
A.3
Comparative statics in table 1
Using the equality ¦E (´¤ ; z) = ¦F (´ ¤; z) ;implicit di¤erentiation yields: d´B dz
@(¦E ¡¦F )
= ¡ @(¦ @z¡¦ E
@´
F)
@M¼ @z = ¡ @M¼ @´
where we can use: 8 < <0:® >® ~ @(¦E ¡¦F ) ; @´ : >0:® <® ~ @(¦E ¡¦F ) @G
A.4
= 1 > 0;
@(¦E ¡¦F ) @t @(¦E ¡¦F ) @®
= ¡s A¡t 2¡´ < 0 2
A ´ = ¡ 12 s (2¡®´) 2 < 0
Appendix: Secondorder conditions
In this appendix, we check the …rm’s secondorder conditions for the maximization of (3). To have a wellposed maximization problem, the Hessian, de…ned in (A.2), must be negative de…nite: 2 3 ¦h;qh ;qh ¦h;qh ;xh 5 Qh = 4 ¦h;xh;qh ¦h;xh;xh where, for example, ¦h;qh;xh =
@ 2¦ h @
[email protected] .
(A.2)
This, in turn, requires that jQhj >
0; ¦h;qh ;qh < 0 and ¦h;xh ;xh < 0. I can show that this will hold if ®´ < 2 since: jQE j = °s (2 ¡ ´); ¦E;qE ;qE = ¡ 2s < 0; ¦E;xE ;xE = ¡° < 0
jQF j = °s ( 2® ¡ ´); ¦F;qF ;qF = ¡ 2s < 0; ¦F;xF ;xF = ¡®° < 0
28
Notes 1See 2
Dunning (1977) and Markusen (1995).
This have been been suggested by Caves (1996).
3Teece
(1977) provides strong evidence for the existence of such technol
ogy transfer costs. 4Brainard
…nds that the elasticity of a¢liate sales with regard to R&D
is 0:1840. The corresponding elasticity for exports is 0:4599. Hence, the a¢liate share of foreign sales should decrease in R&D intensity. 5See 6The
Cheng (1984). assumption of a quadratic transfer cost is not essential. What is
important is that the transfer cost in‡uences the level of R&D, x. 7In
appendix A.4, it is shown that the latter assumption guarantees that
the secondorder condition for the …rm’s maximization of (3) is ful…lled. 8These
conditions are necessary in order to guarantee that production is
pro…table. 9
A decrease in ´¤ is indicated by a minus sign, an increase in ´ ¤ by a
plus sign. 10The
countries included are: Belgium, France, Italy, Holland, Germany,
Luxemburg, UK, Norway, Ireland, Denmark, Spain, Portugal, Greece, Finland, Austria, Switzerland, USA, Canada, Japan, Australia and New Zealand. The last two countries are combined into one single country observation.
29
11
It can be shown that R&D intensity, de…ned as the share of R&D expen
ditures in total sales, is positively correlated with our theoretical measure of return to R&D, ´. 12The
eightyear lag is constructed by lagging R&D two surveys back.
This produces eightyear lags for most observations. The exception is 1986, where the lag is taken from 1974. Similarly, the fouryear lag used in 1986 is taken from the 1978 survey. I have taken this approach because no survey was undertaken in 1982. 13R&D
performed at home completely dominates total R&D expenditures
in the …rms of this sample, even though the share of R&D performed abroad has increased over the period. 14The
reason why size is only used as a control variable is that s is included
in the de…nition of ´, and therefore a¤ects the R&D intensity. But since I aim at capturing the implementation choices of new technologies through AFSHARE, it is still necessary to control for size e¤ects. 15This
type of externalities may involve the use of joint networks of sup
pliers and distributions (see, for example, Venables (1996)). 16There
is a similar pattern in table 4 where a comparison is made with
panel analysis 17
See, for example, Fors (1997).
30
Table 1: Description of variables. Variable name
RDINT
Proxy for
Expected
Description and
sign
source
´
share of a …rm’s total R&D expenditures in its total sales, lagged eight years in sp eci…cation (ii) and (vi) present intensity in sp eci…cations (i), (iii)(v) and (vii). (IUI).
TREMB
t
+
share of transport and packing costs in total variable costs divided by the total industry mean at the three or fourdigit level in the Swedish industry to which the …rm belongs (SCB).
DIST
t
+
distance from Sweden to the respective countries where the …rm records foreign sales. (IUI).
AGE1
®
+
weighted average of the mean age of the …rm’s a¢liates in the respective countries where production takes place. Weights calculated as the share of the …rm’s total foreign sales attributed to the individual countries. (IUI).
RD1
T
+
dummy variable that takes on the value of one if the …rm undertakes any R&D abroad, zero otherwise. (IUI).
GSCALE1
G
¡
average size of the a¢liates divided by the average size of the …rms in terms of employees at the three or fourdigit industry level to
31
which the …rm belongs. (IUI).
Table 2: Continued Variable
Proxy
name
for
AGE2
®
Expected
Description and
sign
source
+
mean age of a …rm’s a¢liates in a speci…c country. (IUI).
RD2
®
+
dummy variable that takes on the value of one if the …rm undertakes any R&D in a country, zero otherwise. (IUI).
GSCALE2
G
¡
average size of plants with more than one hundred employees divided by total industry mean size at the three or fourdigit level Swedish industry, to which the …rm belongs. (SCB).
GDP
PPPadjusted, de‡ated GDP. (OECD, World Bank).
INCOME
ratio between PPPadjusted, de‡ated GDP per capita in Sweden and the respective countries where the …rm records foreign sales. (OECD, World Bank).
OPEN
index measuring the openness of a country to FDI. (Wheeler and Moody (1992)).
AGGLEM
share of total employment in an industry at the three or fourdigit industry level in the respective countries where a …rm records foreign sales. 32 (Braunerhjelm and Svensson (1996) and OECD).
Note: Column two describes the exogenous variable to which the proxy refers. As the theoretical section involves two models with both di¤erent variables and di¤erent predictions, the top row for each exogenous variable corresponds to model 1, whereas the bottom row corresponds to model 2.
Table 3: Twostage Heckman estimation and 2SLS. Variables
RDINT
TREMB
DIST
AGE1
RD1
GSCALE1
Speci…cation (i) Probit
OLS
2SLS
Probit
OLS
2SLS
0.104
0.251
0.333
0.116
0.173
0.236
(4 .107)
(11. 095)
(9 .775)
(3 .274)
(5 .311)
(5 .102)
0.020
0.085
0.097
0.095
0.088
0.114
(0 .386)
(2. 400)
(2 .448)
(1 .120)
(1 .737)
(2 .104)
0.566
0.123
0.381
0.511
0.213
0.419
(11 .544)
(2. 499)
(7 .544)
(8 .238)
(3 .095)
(6 .998)
0.245
0.231
(11 .041)
(5.683)
0.225
0.276
(4 .102)
(3.207)
0.275
0.414
(7 .429)
(7.494)
AGE2
0.114
0.125
0.146
0.126
(7.278)
(5.762)
(5.613)
(4 .948)
0.263
0.254
0.167
0.204
(6.718)
(5.386)
(3.187)
(3 .737)
0.947
0.108
0.213
0.320
(1.787)
(1.466)
(1.989)
(2 .964)
0.347
0.061
0.040
0.390
0.124
0.023
(14.547)
(2.250)
(1 .869)
(12 .290)
(3 .141)
(0 .904)
RD2
GSCALE2
GDP
LAMBDA
Prediction
Speci…cation (ii)
0.545
0.613
(5. 974)
(5 .167)
22 .6
27.9
errors (%) Chi2
817 .68
2
333.43
R
25. 4
23.6
30.4
F
13. 15
11.55
8.92
29 .0 9 .30
No. of var.
26
27
25
24
25
24
No. of obs.
4283
1462
1065
2102
797
775
33
Note 1: The dependent variable in the OLS columns is the a¢liates share of foreign sales for …rm i in country j at time t. The depe ndent variable in the probit columns is a dummy variable which equals one if production is registered, zero otherwise. In the 2SLS re gressions, the eight year lag and the present intensity are respectively used as instruments for the four year lag in R&Dintensity. Note 2: Numbers in parenthesis are tstatistics. Prediction errors are formed at a critical probability of 0.5. All variables are in logs, except RD1, RD2 and LAMBDA. Sample size di¤erences re‡ect missing observations. Intercept and dummies for region, industry and time are not shown for Speci…cations (ii) and (iii).
Table 4: A comparison with panel analysis. Variables
RDINT
TREMB
DIST
AGE1
RD1
GSCALE1
Sp eci…cation (iii)
Speci…cation (iv)
Speci…cation (v)
Probit
OLS
Logit
OLS
Logit
OLS
0.111
0.235
0.311
0.204
0.578
0.112
(4 .163)
(10. 478)
(2 .430)
(4. 327)
(2 .684)
(2 .696)
0.013
0.067
0.211
0.092
0.152
0.129
(0 .259)
(1. 829)
(0 .875)
(1. 128)
(0 .401)
(1 .743)
0.453
0.173
(6 .979)
(3. 406)
0.236
0.203
0.511
(10 .246)
(2 .225)
(3 .432)
0.215
0.293
0.794
(3 .806)
(1 .620)
(2 .517)
0.288
0.257
1.808
(7 .526)
(1 .809)
(4 .233)
AGE2
RD2
GSCALE2
GDP
INCOME
OPEN
AGGLOM
0.126
0.106
0.058
(6. 353)
(5. 385)
(2 .455)
0.255
0.201
0.195
(6. 420)
(3. 976)
(3 .477)
0.075
0.138
0.115
(1. 445)
(2. 341)
(2 .241)
0.319
0.048
1.078
0.166
2.215
0.484
(11.574)
(1.920)
(1.023)
(0. 409)
(1 .080)
(1 .115)
0.011
0.127
0.014
0.122
0.105
0.015
(0.120)
(1.900)
(0.068)
(1.507)
(0 .247)
(0 .174)
0.200
0.009
(1.346)
(0.086)
0.139
0.181
0.216
0.129
0.517
0.282
(2.208)
(3.487)
(1.694)
2.137)
(0 .903)
(1 .673)
LAMBDA
0.553 (5. 885)
Prediction
23 .7
34 34 .5
39 .4
699 .90
635 .65
102 .90
errors (%) Chi2
R
2
26. 6
20. 9
11. 58
F
0 .2
9. 39
3 .98
No. of var.
28
29
32
32
14
15
No. of obs.
3811
1343
3716
1285
696
1040
177
141
186
319
No. groups
Table 5: Comparative statics results in model 1 Low transfer costs, ® > ® ~: (FDI when Variable:
E¤ect on
´¤
´ > ´¤ ) ME on FDI
High transfer costs, ® < ® ~: (FDI when
´ < ´ ¤)
E¤ect on ´ ¤
ME on FDI
G
(+)

()

®
()
+
(+)
+
t
()
+
(+)
+
Note: Note that an increase in alfa implies a decrease in technology transfer costs.
35