Estimating the Knowledge-Capital Model of the ...

Estimating the Knowledge-Capital Model of the MultinationalEnterprise: Comment

By BRUCE A. BLONIGEN, RONALD B . DAVIES, AND KEITH HEAD*

A recent American Economic Review articleby David L. Carr et al. (2001) (CMM) estimatesa regression specification based upon the"knowledge-capital" model of the MultinationalEnterprise (MNE). The knowledge-capitalmodel combines "horizontal" motivations forforeign direct investment (FD!)—the desire toplace production close to customers and therebyavoid trade costs—with "vertical" motiva-tions—the desire to carry out unskilled-labor-intensive production activities in locations withrelatively abundant unskilled labor. By way ofcontrast, the horizontal model, an intellectualantecedent of the knowledge capital model, pre-cludes the separation of knowledge-generatingactivities from production and therefore gener-ates different policy implications. CMM's sum-mary states that the results "fit well with the[knowledge-capitall theory. We hope that themodel will therefore prove useful in future anal-ysis." In this Comment, we argue that ratherthan offering direct support for the knowledge-capital model, the data set used by CMM cannotreject the horizontal model of MNEs in favor ofthe knowledge-capital model.

The crux of the distinction between theknowledge-capital model and the horizontalmodel lies with the estimate of the eflect of skilldifferences on the level of affiliate activity inthe host country. CMM find that the increases inthe parent-country's relative skill endowmentrai.se affiliate sales in the host country as long as

* Blonigen: Department of Economics, t285 LIniversityot" Oregon. Eugene. OR 97403 (e-mail: bmceb@Oregon,uoregon.edu); Davies: Departmeni ol' Economics. 1285Universily of Oregon, Eugene. OR 97403 (e-mail: rdaviestfporegon.uoregon.edu): Head: Faculty of Ciimmerce. Univer-sity of British Columbia, 205.'? Main Mail. Vaneouver, BCV6T1Z2. Canada (e-mail: [email protected]). We wouldlike to thank Robert Feenstra. Stephen Haynes. JamesMarkusen. Keith Ma,skus. Kaoru Natwshima, Matt Slaughter,Jim Ziliak. and participants of a 2002 American EconomicsAssociation session tbr helpful discussions and comments.We also thank Sarah Lawson for excellent research assis-tance. Any remaining eiTors or omissions are our own.

the parent country is small. However, this effectof skill differences is decreasing in the parent-host GDP difference. They interpret these re-sults as support for the knowledge-capitalmodel of MNEs. We demonstrate that this find-ing arises because of a misspecification of theskill difference terms in their empirical frame-work. When corrected, we find that absoluteskill differences reduce affiliate sales. This in-stead supports the horizontal model of the MNEand suggests that it cannot be rejected in favorof the knowledge-capital model. Our findingsare robust to alternative specifications usingboth U.S. and OECD data.

Interest in MNEs has grown considerably inrecent years for two main reasons. First, flows ofFDI. the defining activity of MNEs, have grownat substantial rates over the last two decades,outstripping the rate of growth of both worldoutput and intemational trade. Second, there hasbeen an increasingly vocal public and academicdebate on the effects of FDI, particularly withrespect to labor market effects. This debate hasbeen informed by several models of FDI, par-ticularly those of James R. Markusen and coau-thors.' These models are especially relevant tothe debate on FDI and wages because theysuggest many different motives for engaging inFDI and thus many different potential labormarket effects. For example, affiliate activity inforeign countries is less likely to have a nega-tive impact on unskilled home-country workersin a horizontal modei than a knowledge-capitalmodel.

As with the literature testing models of inter-national trade, researchers have turned to thedata to select the most appropriate model of theMNE. CMM develops an empirical frameworkto study the efficacy of the knowledge-capital

'These include Markusen (I9S4, 1997); Ignatius J.Horstmann and Markusen (1987, 1992): Markusen et al.(1996); and Markusen and Anthony S. Venables (1997.1998. 2000).

980

VOL. 93 NO. 3 BLONIGEN ET AL: THE MULTINATIONAL ENTERPRISE. COMMENT 981

model of MNE activity. The CMM estimatespool inward and outward U.S. affiliate salesdata from 1986 through 1994 and appear tosupport the knowledge-capital model of theMNE. In particular, the terms they use as prox-ies for skilled-labor-abundance differences be-tween countries, the key variables identifyingvertical MNE motivations, have the expectedsigns and are statistically significant. However, inrelated work, Markusen and Maskus (1999. 2001)find evidence confiicting with the knowledge-capital model using the same database.

In this Comment on CMM we resolve thisapparent puzzle. We show that CMM's empir-ical framework niisspecifies the terms measur-ing differences in skilled-labor abundance. Aftercorrecting this specification error, the coefficientestimates no longer support the knowledge-capitalmodel. Instead, the data strongly support thepredictions of the horizontal model of MNEs:affiliate activity between countries decreases asabsolute differences in skilled-labor abundancewiden. Further, we strengthen the evidence forthis result by showing that the negative relation-ship between FDI activity and dissiinilarity inskilled-labor abundance is also found using datathat include a wider variety of parent and hostcountries, including data for the OECD. Finally.we show exactly how the difference betweenCMM and Markusen and Maskus (1999, 2001)follows directly from the misspecification ofskill differences.

This paper proceeds as follows. In Section I,we briefly summarize the varied theories of FDIand survey the small empirical literature on thedeterminants of FDI. In this same section, wedetail the specification error in the CMM empiri-cal framework used to estimate the knowledge-capital model. Section II uses the CMM data setto illustrate the stark change in coefficient esti-mates when we correct the specification errorand shows that the same coefficient pattemsappear in altemative U.S. and OECD samples ofMNE activity. We do not find support in any ofthese data sets for rejecting the horizontal modelin favor of the knowledge-capital model. Sec-tion III concludes.

I. Recent Evidenee on MNE Models: A Puzzle

Relative to many prior empirical studies ofFDI, the CMM approach represents a step for-

ward because it bases its framework in the for-mal theories of the multinational firm. Thesetheories can be divided into three rough catego-ries: the horizontal model, the vertical model,and the knowledge-capital model. The horizon-tal model originates in Markusen (1984) anddescribes a firm with plants that engage in thesame activity in multiple locations. This modelposits that FDI arises from an interaction be-tween firm-level economies of scale and tradecosts. Markusen and Venables (2000) showthat, in the horizontal model, dissimilarity inrelative endowments reduce the activity ofMNEs; thus the horizontal model predicts thatabsolute skill differences should be negativelyrelated to FD! activity.

The vertical model, first formalized by El-hanan Heipman (1984), builds an incentive tolocate different activities in different countriesin order to take advantage of factor cost differ-ences. One strong prediction of this model isthat FDI should only flow from the skill-abundant country to the unskilled country (sincea firm's nationality is identified with the loca-tion of its skill-intensive headquarters). Further-more, when countries are identical, there is noreason to engage in FDI since there are no costdifferences to exploit.

More recently, Markusen et al. (1996) andMarkusen (1997) have developed the knowledge-capital model tested in CMM. This model inte-grates lhe horizontal and vertical models andallows for both multiplant scale economies andexploitation of factor-price differences. Sincethe knowledge-capital model is a combinationof the horizontal and vertical models, it comesas no surprise that skill differences can havepositive and negative effects. Specifically, a risein skilled-labor-abundance differences tends toincrease FDI from the skilled country to the host(as predicted by the vertical model). This effectdiminishes, however, when the unskilled host issmall. Thus, while the total effect of skill dif-ferences is ambiguous due to the interactionswith country size, a more skilled large parentcountry should have more outbound FDI than aless skilled or a small parent country.

A key distinction between the pure horizontalmodel and the knowledge-capital model is thatthe former assumes that headquarter servicesand production use factors in the same propor-tions. In contrast, the knowledge-capital model

982 THE AMERICAN ECONOMIC REVIEW JUNE 2003

assumes the headquarter services use skilled labormore intensively than all other activities. Never-theless, there are regimes in the knowledge-capitalmodel where multinationals of the horizontalform (plants in more than one country, head-quarter services in only one) do emerge. Thisoccurs when relative endowments are similarsince this removes the factor-price differencesthat generate the incentives for the verticalform. Thus, observing MNE activity betweencountries of similar factor proportions does notviolate the knowledge-capital model. Neverthe-less, the distinguishing feature is that a diver-gence in relative factor endowments reducesproduction of foreign affiliates in both countriesin the horizontal model, but can increase it inthe knowledge-capital model.

With three altemative models of MNE activ-ity, empirical investigation naturally followed,including a set of papers by Markusen andMaskus: Markusen and Maskus (1999),Markusen and Maskus (2001), and CMM(2001).^ All three use data on U.S. affiliate salesabroad (outbound affiliate sales) and sales offoreign affiliates in the United States (inboundaffiliate sales) from 1986 through 1994 to in-vestigate the various models of MNE activity.All three papers motivate reduced-form empir-ical specifications from simulated topologies ofMNE activity over alternative variable andparameter spaces derived from the general-equilibrium modeling of Markusen in previoustheory work. The topologies are often nonlin-ear, leading to interaction and squared termsin the empirical specification, in particular,CMM pools observations of both inbound andoutbound U.S. affiliate sales, tests it on atheory-motivated empirical specification ofthe knowledge-capital model, and finds seem-ingly robust support for this MNE model.Markusen and Maskus (2001) extend CMM'swork by exploring additional empirical impli-cations of the knowledge-capital model and es-

^S. Lael Brainard (1997| develops a horizontal MNEmodel where firms are located to foreign markets tbr "prox-imity advantages" and finds evidence consistent with lhehori/,ontal model using U.S. liata. Karolina Ekholm (1995,1997, 1998a, b) empirically examines implications ofMarkusen's knowledge-capital model, but does nol try loconnect il as direcdy to the theory as in lhe three papers wediscuss in thin section.

timating whether these exist in the same data setof inbound and outbound U.S. affiliate activity.While that paper still concludes that there issubstantial evidence for the knowledge-capitalmodel, they find a surprising result that skilled-labor-abundance differences (parent minus host)are significantly negatively related to EDI ac-tivity in the outbound U.S. data. Eurthermore,the coefficient on skill differences interactedwith GDP differences is positive and signifi-cant. These results conflict with the predictionsCMM provide for the kjiow I edge-capital model.

Markusen and Maskus (2001) suggest thatthe difference with CMM could derive from twosources. Eirst, the theory models a two-countryworld where total world endowments are fixed,whereas their data set contains observations oncountry-pair observations with vEU"ying endow-ment totals. Since it is unclear what impact thismight have on the model's predictions, the re-sults for outbound affiliate activity in Markusenand Maskus (2001) may not contradict theknowledge-capital model. Second, they notethat it may be problematic that the United Statesis one of the two countries in every country-pairobservation in their sample. Since the UnitedStates is substantially larger than every othercountry, this restricts the observations to only acertain region of the parameter space, whichcould then skew the empirical results.

Einally, Markusen and Maskus (1999) use thesame database of U.S. inbound and outboundaffiliate activity as in CMM to examine an em-pirical specification that they present as nestingall three models of MNE activity: horizontal,vertical, and knowledge-capital. Their empiricalframework differs from that found in CMM.Whereas CMM specified skill differences as thedifference between the skilled-labor abundanceof the parent to the host country, Markusen andMaskus (1999) include additional interactionterms that indicate when this relative skill dif-ference between the parent and host country ispositive versus when it is negative: i.e., whenthe parent country is relatively ski lied-labor-abundant versus when the parent country isskilled-labor-deficient. As we explain and dem-onstrate in this paper, this distinction betweenregions when the skill difference is negativeversus vfhen it is positive is critical. The empir-ical evidence in Markusen and Maskus (1999)strongly supports the horizontal model, and re-

VOL 93 NO. 3 BLONIGEN ET AL: THE MULTINATIONAL ENTERPRISE. COMMENT 983

jects the vertical and knowledge-capital models,which contrasts sharply with the conclusion ofCMM.

Taken together, the recent evidence on MNEmodels presents a puzzle. Since all three papersuse the same database on U.S. inbound andoutbound MNE activity and are derived fromthe knowledge-capital model, why do they pro-duce contradictory results regarding skilldifferences?

As the effect of differences between the par-ent and host country skill endowments is themajor distinction between the predictions of theknowledge-capital and horizontal models, theresolution of the puzzle rests with this issue.The key is to realize that interpretation of thecoefficients on such a difference variable de-pend critically on whether the sign of the dif-ference term is negative or positive. When theskill difference term lies in the positive range,an increase in the variable corresponds to agreater inequality in relative skill endowments.However, when the skill difference term is neg-ative, parent- and host-country skill endow-ments converge as the difference term rises. Asa result, it is incorrect to estimate a pooledcoefficient on a difference term that takes bothpositive and negative values in the sample. Asshown by Stephen E. Haynes and Joe A. Stone(1981), difference terms impose a subtractivelinear constraint which can lead to a sign rever-sal in the pooled (or restricted) coefficient.Haynes and Stone show that this sign reversalindeed occurs in the estimation of a real interestrate differential model of the exchange rate byJeffrey Erankel (1979). and we likewise showthe same problem of sign reversal occurs inCMM.

Eor the bilateral U.S. affiliate data used by thepapers discussed above, the skill differenceterm lies predominantly in the positive regionfor U.S. outbound affiliate activity and predom-inantly in the negative range for the inboundaffiliate activity in the United States. With bothinbound and outbound affiliate sales pooled intoone sample as in CMM, it is difficult to interpretthe single coefficient on the skill difference termsince it takes both positive and negative values.If the knowledge-capital model is correct andone separates out the skill difference terms intothose observations where it is in the positiveregion (i.e., outbound U.S. affiliate activity) and

the negative region (i.e., inbound activity), wewould expect the same coefficient signs in bothregions. In contrast, the horizontal model pre-dicts opposing signs.

In fact, Markusen and Maskus (2001) obtain"correct" signs for the knowledge-capital modelin the inbound sample, whereas the outboundsample has the reverse signs. Thus EDI activitydecreases when absolute skill differences rise.Even more convincingly, Markusen andMaskus (1999) specifically take into accountthe expected sign reversal by interacting theskill difference with a dummy variable indicat-ing whether the difference term is in the nega-tive or positive region. They conclude that sincethe signs in their empirical analysis are exactlythe opposite of those predicted by the knowledge-capital model in CMM, the data supports thehorizontal model. In their discussion of thisdiscrepancy between the two papers, they donot attribute the resolution of the puzzle to theissue of whether the difference term is generallyin the negative region or the positive region.

II. Results

To examine our proposed resolution, we ob-tained the data used in CMM from the authors.We were able to exactly replicate their resultsfor all their reported specifications. Columns 1and 3 of our Table 1 show coefficient estimatesfor the knowledge-capital model using ordinaryleast squares (OLS) and Tobit specificationsthat correspond exactly to those CMM report intheir Table 3.

We then modify their framework into whatwe term an absolute value model where wespecify the skill difference and GDP differenceterms as absolute values. Specification of thedifference terms in absolute values implies thatthe new variable is always increasing in skilldissimilarity. This facilitates interpretation ofcoefficients and marginal effects of these re-gressors. Columns 2 and 4 of Table 1 reportcoefficient estimates for our absolute value ver-sion of the knowledge-capital model for theOLS and Tobit specifications, respectively. Theresults are striking. The coefficients on both theabsolute skill difference term and its interactionwith the absolute GDP difference are statisti-cally significant at the l-percent significancelevel and of opposite signs to those of CMM.

9S4 THE AMERICAN ECONOMIC REVIEW JUNE 2003

TAHLE 1—OLS RESULTS USING CMM MODEL VERSUS RESULTS FROM ABSOLUTE DIFFERENCE

VERSION OF KNOWLEDGH-CAPJTAL MODEL

Regressors

GDP Sum

GDP Difference Squared

Skill Difference ( = Skillp - Skillh)

Skill Difference * GDP difference( = (Skiilp - Skill,,) * (GDP^ - GDP,,))

Absolute Skill Difference (= Skillp -SkiilJ)

Absolute Skill Difference * Absolute GDPDifference ( = |Skillp - Skilly * |GDPp- GDPJ)

Inveslment Cost Host

Trade Cost Host

Trade Cost Host * Squared SkillDifference

Trade Cost Parent

Distance

Intercept

ObservationsAdjusted ff'Log-likelihood

Table 3 resultsfrom CMM

10.80**(7.01)

-0.0012**(-6.89)

33,743*^^(3.77)

-6.34**(-2.62)

-516.6**(-3.79)

119.2(1.16)605.2(0.36)-93.7

(-0.99)-1.82**(-7.75)16.630(1.08)5090.46

OLS

Absolute differencemodel

17.57**(12.13)

-0.0040**(-14.77)

-1.485,525**(-12.85)253.39**(12.08)

-173.1(-1.75)-109.2(-1.08)5.997**(2.84)

-108.6(-1.32)-1.31**(-6.34)

57.437**(4.18)5090.59

Table 3 resultsfrom CMM

15.04**(10.27)

-0.0010**(-5.89)

61,700**(7.28)

-10.20**(-4.34)

-387.6**(-2.82)

156.2(1.51)

-1.264(-0.75)-122.0(-1.46)-1.48**(-6.47)-23.283(-1.61)

628

-5.755

Tobil

Absolute differencemodel

21.24**(15.02)

-0.0037**(-13.19)

-1,428.759**(-12.07)234.21**(10.87)

229.8(2.39)

-227.3*(-2.22)6,896**(3.15)

-201.2**(-2.75)-1.00**(-4.86)17,300(1.29)628

-5.716

Note: p = parent, h = host, /-stati.stics are in piirentheses. with *"'the 1- and 5-percent levels, respectively.

and * cienoiing statistical significance (two-tailed test) at

The independent effect of skill differences nowstrongly suggests that real affiliate sales de-crease as skill levels diverge—the opposite re-sult from the sign predicted by CMM for theknowledge-capital model. Likewise, the inter-action term of skill difference with GDP differ-ence has the opposite sign and statisticallysignificant at the l-percent level when the dif-ference terms are specified in absolute values.This, too, conflicts with the predicted sign of theknowledge-capital model.^ Finally, the R^ risessubstantially in the OLS model when using theabsolute value model (from 0.46 to 0.59) and

•' The other right-hand side (RHS) term in which skilldifference enters in the CMM empirical framework is theinteraction of the squured Skill Difference wilh Trade Co.stHost. Since it is squared, no absolute value correction isneeded for this term.

the value of the log-likelihood increases for theTobit specification (from —5,755 to —5,716).The original CMM aiticle also presented aweighted least-squares (WLS) specification, aswell as host-country fixed-effects versions ofthe OLS, WLS, and Tobit specifications. Weobtain identical changes in the signs of the skilldifference regressors when using the absolutevalue model that are statistically significant atthe l-percent level for all of these additionalspecifications (these results are available uponrequest). In summary, once these skill differ-ence terms are appropriately specified, the dataoffer no support for the predictions of theknowledge-capital model with respect to skilldifferences between countries.

The use of an absolute difference model,however, still involves a restriction that skilldifferences have symmetric effects on real affil-

VOL. 93 NO. 3 BLONtGEN ETAL: THE MULTINATIONAL ENTERPRISE, COMMENT 985

TABLE 2—OLS RESULTS USING CMM FRAMEWORK TO COMPARE EsTiMArEs FROM SAMPLE WITH POSITIVE SKILL

DIFFERENCES VERSUS SAMPLE WITH NEGATIVE SKILL DIFFF.KENCF.S AND SAMPLE OF OUTBOUND U.S. AFFILIATE SALES

VERSUS SAMPLE OF INBOUND U.S. AFRLIATE SALES

RegressorsSample with positive

skill differencesSample with negative

skill differences Outbound sample Inbound sample

GDP Sum

GDP Difference Squared

Skill Difference (=Skillp - SkiU^)

Skill Difference * GDP Difference( = (Skill,, - Skill,,) * (GDPp -GDP,,))

Investment Cost Host

Trade CosI Host

Trade Cost Host * Squared SkillDifference

Trade Cost Parent

Distance

Intercept

ObservationsAdjusted R^

9.21**(5.86)

-0.0014**(-7.5)

-81.147**(-2.83)

5.33(1.50)

-522.3**(-3.96)

52.3(0.47)3,863(1.82)

-420.5*(-2.46)-1.59**(-6.25)

46.694**(2.86)3060.54

13.14**(4.04)

-0.0012**(-3.56)

220,693**(2.66)

-10.82(-1.49)

-1,097(-1.37)-46.3

(-0.12)27,825*(2.50)-73.0

(-0.52)-2.21**(-5.14)34,589(0.88)2030.45

15.50**(9.16)

-0.0045**(-15.08)

-1,575.770**(-13.02)289.95 *=i(12.87)

-1 .031**(-7.88)417.8**(4.37)540.6(0.33)123.4(0.52)

-2.17**(-8.16)

92.987**(5.56)3100.67

22.37**(8.18)

-0.0024**(-7.10)

989,126**(7.78)

172.4**(7.34)

389.2(0.36)-39.7

(-0.12)-3.047(-1.11)-2.80

(-0.03)-0.73**(-2.67)-52,763(-1.23)

1990.64

Nate: p = parent, h = host, /-statistics are in parentheses, wilhthe 1- and 5-percent levels, respectively.

and * denoting statistical significance (two-tailed tesl) at

iate sales. In other words, the relationship be-tween absolute skill differences and real affiliatesales is identical for instances where the parentcountry is more skilled-abundant compared tothe host (skill difference term is positive invalue), as well as where the host country ismore skilled-abundant compared to the parent(skill difference term is negative in value). Inthe U.S. data used by CMM. the former instanceof positive skill differences is almost entirelyobservations of U.S. affiliate sales abroad (out-bound affiliate sales), and the latter instance isalmost entirely observations of foreign affiliatesales in the United States (inbound affiliatesales).

In Columns 1 and 2 of Table 2, we split theCMM sample into observations where the skilldifference is always positive and where the skilldifference term is always negative. If diver-gence in skill levels leads to a symmetric de-cline in real affiliate sales, then we shouldexpect a negative coefficient on the skill differ-ence term in the sample of positive skill differ-

ences and a positive coefficient for the sampleof negative skill differences and they should beof comparable magnitude. The signs of coeffi-cients in columns 1 and 2 of Table 2 confirm thesign expectation and again support the horizon-tal model predictions. The magnitudes, how-ever, are not of equal size, with the coefficienton skill difference for the sample of negativeskill differences almost three times as large. Aswe will discuss below, the marginal effects ofskill differences on real affiliate sales (whichalso takes into account the interactions of skilldifference with GDP difference and host-country trade costs) exhibit a similar change inmagnitude.

For comparison, columns 3 and 4 of Table2 provide estimates for separate samples of out-bound and inbound affiliate activity. The coef-ficients on the skill difference terms likewiseshow opposite signs across the sample indicat-ing that real affiliate sales and absolute skill dif-ferences are negatively related, in contrast withCMM's predictions for the knowledge-capital

986 THE AMERICAN ECONOMIC REVIEW JUNE 2003

model. For the outbound-inbound split, the skilldifference coefficient is now larger for the out-bound sample, not the inbound sample. This isseemingly inconsistent with the positive-nega-tive sample split. However, in the marginal ef-fects reported below, once one takes intoaccount the interaction terms, the relative ef-fects of skill difference on affiliate activity arequalitatively identical: an increase in skill dif-ference leads to a decrease in affiliate activitythat is approximately three times larger for theinbound (negative skill difference) sample asthe outbound (positive skill difference) sample.

It is important to note that the opposite coef-ficients for the inbound and outbound samplecorrespond to the results found by Markusenand Maskus (2001). It is clear these opposingcoefficient estimates for the two samples comefrom the fact that the skill difference is primar-ily negative in value in the inbound sample andpositive in the outbound sample.

A. Marginal Effects

Given the interaction terms involving theskill difference term in the CMM framework,the coefficient estimate on the skill differenceterm is not the marginal effect. For the OLS andWLS regressions, the marginal effect is

3(Real Affiliate Sales)/<5(Skill Difference)

= B3 + B4(GDP Difference)

+ 2 * B7 * (Trade Cost Host)

* (Skill Difference),

where B3 is the estimated coefficient on theskill difference term, B4 is the estimated coef-ficient on the interaction between skill differ-ence and GDP difference, and B7 is theestimated coefficient on the interaction betweenskill difference squared and host-country tradecosts .' In Table 3 we report marginal effects fora standard deviation change in skill differencesfor our absolute value models in Table 1 and our

"* Marginal effects of the Tobit specifications must alsotake into account the truncation of the sample, which wasapparently not done in CMM—see Panel E of Table 2.

sample splits in Table 2. The marginal effects inevery case correspond to the estimated coeffi-cient on the skill difference term by itself. Inother words, the interaction ternis are not coun-teracting the estimated independent effect ofskill differences, and our inferences that skilldifferences are inversely related to real salesactivity are confinned and statistically significant.

The marginal effects for the specificationsthat correctly model skill differences all suggestan inverse relationship between skill differencesand real affiliate sales activity that is substantialin magnitude. For example, at the means of thedata, a standard deviation increase in the abso-lute skill difference (a change in the share of acountry's skilled worker share of approximately10 percentage points) reduces real affiliate salesby $7.1 billion in the OLS absolute valuemodel, where the average real affiliate sales inthe sample is $15.8 billion. When the sample issplit into inbound and outbound activity, theeconomic effect of skill differences on real af-filiate sales is revealed to be much more pro-nounced for inbound activity, where an increasein skill dissimilarity (a negative change in theskill difference term for all observations that arenegative) leads to an $8.0-billion decrease inreal affiliate sales. This effect is over three timeslarger than the $2.3 billion decrease on theoutbound side for a standard deviation increasein skill differences.

Marginal effects of skill difference on realaffiliate sales are also calculated and discussedin Result 4 of CMM. Rather than calculating asingle marginal effect at the means of the data,they calculate marginal effects for every bilat-eral pairing in the sample for the year 1991 andthen report separate marginal effects for in-bound and outbound observations. The signs oftheir marginal effects for inbound and outboundobservations agree with the results we reporthere (real affiliate sales negatively related toabsolute skill differences), even though theircoefficient estimates on the skill differenceterms are completely the opposite of ours. Thereason CMM obtain negative marginal effectsin the outbound sample despite a positive coef-ficient on the skill difference term (B3) is thepresence of the negative coefficient on the in-teraction between skill and GDP differences(B4). As the United States has a much higherGDP than the countries it invests in, the out-

VOL 93 NO. 3 BLONIGEN ET AL: THE MULTINATIONAL ENTERPRISE, COMMENT 987

TABLE 3—MARGINAL EFFECTS OF SKILL DIFFERENCES ON REAL AFFILIATE SALES/STOCK EVALUATED

AT THE MEANS OF THE DATA

Model

CMM SampleCMM OLS Difference Model (Column

1 of Table 1)OLS Absolute Difference Model

(Column 2 of Table 1)CMM Tobit Difference Model (Column

3 of Table 1)Tobit Absolute Difference Model

(Column 4 of Table 1)OLS Difference Model, Positive Sample

(Column I of Table 2)OLS Difference Model, Negative

Sample (Column 2 of Table 2)OLS Difference Model. Outbound

Sample (Column 3 of Table 2)OLS Difference Model, Inbound Sample

(Column 4 of Table 2|Modified U.S. SampleOLS Absolute Difference Model

(Column 2 of Table 4)OECD SampleOLS Absolute Difference Model

(Column 4 of Table 4)

Change in real affiliate sales/stockfor a standard deviation change in

skill differences

3,308

-7.137

2.981

-6.531

-2.119

6.576

-2.329

7,975

-15,966

-3.361

f-statistic(p-value)

10.66(0.000)65.02

(0.000)53.88

(0.000)127.25(0.000)

3.68(0.056)10.77

(0.001)3.96

(0.048)50.03

(0.000)

216.6(0.000)

90.92(0.000)

Sample average realaffiliate sales/stock

15,767

15,767

12,779

12,779

14.589

17,542

15.942

15,494

20,821

4.322

Notes: All marginal elTecis are calculated at the means of the data. Real affiliate sales (real stock for OECD) are in millionsof real U.S. dollars. Wald statistic, not F-slatistic, was calculated tbr Tobit marginal effects.

bound sample's marginal effects are dominatedby the GDP difference interaction term. Whiletheir marginal effects correspond in sign to ours,their miuginal effects are still based on coefficientestimates that incorrectly pool inbound and out-bound obsei'vations.' As a result, we find that theirmarginal effects strongly understate the relation-ship between skill differences and affiliate activityfor any comparable calculation.^

^ In other words, marginal effects depend on (1) thecoefficient estimates, and (2) the value of the viiriables usedlor the interaction lemis: skill difference. GDP difference,and host trade cost. By calculating marginal effects at valuesof the data ihat corresponded to various inbound and out-bound observations, CMM got closer to properly adjustingfor the problem with the skill and GDP difference terms, buttheir marginal effects still use pooled coefficient estimatesthat did not account for this problem.

'' Table 3 obviously compares marginal effects calcu-lated at the means of Ihe data. Another comparison can bemade by calculating the miirginal effects ibr every obser-vation in the sample for both studies and then comparing thevalue of the median marginal effect from each sample. Theirmedian marginal effect suggests that a standard deviation

In the end, our estimates suggest a significantnegative relationship between skill dissimilarityand real affiliate sales for both inbound andoutbound MNE activity. A simple scatterplot ofthe data also provides persuasive visual evi-dence of this relationship. Figure 1 plots realaffiliate sales (averaged from 1986 to 1994)activity versus the skill difference term. MNEaffiliate sales are largest when skill differencesare closest to zero and decline as skill differ-ences increase in either direction. While notperfectly comparable. Figure I suggests a the-oretical relationship that is much more in linewith figures of horizontal MNE activity inMarkusen and Maskus (1999). Those figuresdiffer from Figure 2 of CMM which depictsaffiliate sales for a parent country in an Edge-worth box where the skill difference term getsmore positive as one moves northwest from the

increase in skill difterences leads to a $l.04-billion fall in realaflitiate sales, whereas ours suggests a $t9.85-bi!lion fall.

988 THE AMERICAN ECONOMIC REVIEW JUNE 2003

-0.2 -O.I 11,0 O.I 0.2

Parent-Host Skill Difference

FIGURE 1. AFFILIATH SALES (1986-1994 AVERAGES)

diagonal and more negative as one movessoutheast off the diagonal, interestingly, Eigure2 of CMM is consistent with an inverse rela-tionship between skill differences and affiliatesales for negative skill differences between theparent and host, but the same figure also showslittle to no MNE activity unless the parent is nottoo skill-deficient from the host country. Predic-tions of little MNE activity for a skill-deficientparent are also displayed for the horizontalmodel in Markusen and Maskus (1999). Thiscontrasts with the U.S. data that shows surpris-ing inbound MNE activity from parent coun-tries that are substantially skill-deficient to theUnited States. Of course, the figures of affiliateactivity created in CMM and Markusen andMaskus (1999) are for certain fixed-parametervalues, so it is not clear whether reasonableparameter values would yield figures that cor-respond more closely with the data.

B. Exploring Alternative Samples

The CMM data cover a relatively small groupof countries and years. Additionally, every ob-servation of a bilateral country-pair includes theUnited States, causing concern that the variationin the data is confined to only particular param-eter spaces. For example, there are very fewexamples in the data of parent countries that aresmaller than the host country, but which have apositive skill difference.^ The main reasons why

' A couple of exceptions are some of the Europeancountries which are indicated as relatively skilled-labor-

the CMM data are limited in country and yearcoverage are data availability for their trade andinvestment cost variables and, to some extent,their proxy for skilled labor.

As a robustness check we constructed twoaltemative samples to estimate the knowledge-capital model. The first is what we call themodified U.S. sample. We make a number ofchanges to the variables used as proxies andexpand the U.S. sample in the process. In par-ticular, we use data on average educational at-tainment by country and year as a proxy forskilled-labor abundance, rather than the share oflabor listed in skilled occupations that was usedby CMM. Second, we tum to alternative tradeand investment cost data that allowed greatercoverage. Third, we used Penn World Tabledata for real GDP." Einally we use affiliate salesin all industries whereas CMM use sales in themanufacturing sector only (this feature of theirdata, not noted in the paper, results in a reduc-tion in the sample of available countries becauseof disclosure issues). The combined effect ofthese changes is a sample comprising over 50percent more observations covering a slightlydifferent set of countries, over somewhat differ-ent years, with quite different proxies for ourvariables. The Appendix has details on datasources, variable construction, and descriptivestatistics.

Columns I and 2 of Table 4 show that despitethe many changes in the sample and variableconstruction, we obtain the same OLS resultsusing this modified U.S. sample as we did inTable 1 using the CMM data sample. Estima-tion of the CMM framework with simple skilland GDP difference terms yields a relativelysmall positive coefficient on the skill differenceterm, while the absolute value model yields amuch larger negative coefficient, which contra-dicts the predicted signs of the knowledge-capital model. As reported in Table 3. themarginal effect for the absolute value model isalso statistically significant and suggests astrong inverse relationship between skill differ-ences and real affiliate sales. Running separate

abundant to the United States in the data. But even in theseca.ses lhe .skill differences are very small.

'^This .step actually limits the sample lo years through1992, rather than 1994, as in CMM.

VOL 93 NO. J BLONIGEN ET AL: THE MULTINATIONAL ENTERPRISE. COMMENT

XABI..E 4—OLS RESU1-T.S U.S. SAMPt>: AND OECD SAMPLE

Regressors

GDP Sum

GDP Difference Squared

Skill Difference ( = Skill|, - Skilly)

Skill Difference * GDP Difference( = (SkiII,, - Skin,,) * (GDPp - GDPh))

Absolute Skill Difference (=|Skillp -SkiilJ)

Absolute Skill Difference * Absoluic GDPDifference (= Skillp - Skill,. *|GDPp -GDP,,|)

Inve.slment Cost Host

Trade Cost Host

Trade Cost Host + Squared SkillDifference

Trade Cost Parent

Distance

Intercept

ObservationsAdjusted R~

MoJitici U.S. sample

Difference Absolute differencemodel

34.38*^"(15.58)

-0.0035**(-9.06)1.859**(4.46)-0.24

(-0.97)

-763.7**(-4.87)

68.9(1.80)

-5.18**(-3.61)

18.31(27.09)

-3.09^^*(-6.61)

-26,122*(-2.01)

7780.52

modei

30.8!"^*(16.58)

-0.0107**(-21.36)

-58,510**(-18.82)12.34**(16.77)

143.3(1.93)

- 172.4**(-4.84)4.66**(3.18)-16.5

(-0.73)-2.13**(-5.42)

104.349**(8.00)7780.67

OECD

Difference moiJel

9.28**(25.88)

-0.0007**(-7.22)272.5**(2.56)

-0.69**(-10,76)

-46.2*(-2.27)-4.14

(-0.92)-1.38**(-4.78)-69.9**(-1.46)-0.25**(-5.74)726.5(0.77)2.460

sample

Absolute differencemodel

9.26**(26.25)

-O.()O(M**(-4.511

-141.5(-0.78)-0.97**(-12.27)

-5.58(-0.29)

-15.93**(-3.56)-0.30

(-0.96)-55.1**(-4.88)-0.24**(-5.49)-443.6(-0.50)2,4600.39

Note: p = parent, h = host, (-statistics are in parentheses, with ** and * denoting statistical significance (Iwo-tailed lest) atthe 1- and 5-percent levels, respectively.

samples of inbound and oulbound observations(or positive and negative skill difference obser-vations) for this modified U.S. sample againsuggests, as with tbe CMM sample, tbat theinverse relationship between absolute skill dif-ferences and real affiliate sales is much moresubstantial for inbound affiliate activity (orwbere parent-host skill differences are nega-tive). These results are available on request.

The .second alternative we consider is a sam-ple of FDI activity involving OECD countries,which helps to alleviate the problems caused byusing a sample where one of tbe countries inevery bilateral-pair observation is the UnitedStates. We collected OECD data on FD! stock,since data on affiliate sales for countries otherthan tbe United States are generally not col-lected, and matched these data with tbe proxyvariables for GDP, skill, and trade/investmentcosts used in our modified U.S. sample. This

created a sample of 2,460 covering 15 OECDparent countries and 38 bost countries (someOECD and some non-OECD) over tbe years1982 through 1992. Again, tbis is an importantsample to consider because it considers a muchbroader range of possible bilateral pairings thanwben one is confined to data wbere the UnitedStates is necessarily one of the countries in thebilateral pair. In addition, the constructed dataset includes almost 2.500 observations^a muchlarger sample than either of the two U.S. sam-ples we use. The trade-off is that the data arelikely not as accurate, with greater measurementerror with the FDI stock data proxying for af-tiliate activity and measurement consistencyproblems across countries. In the Appendix, wediscuss these issues further, as well as details ondata sources, variable construction, atid descrip-tive Statistics.

Columns 3 and 4 of Table 4 report OLS

990 THE AMERICAN ECONOMIC REVIEW JUNE 2003

results for the CMM difference model and ourabsolute difference tnodel for this OECD sam-ple. The results are consistent witb our findingsusing U.S. data. The CMM difference mode!gives a pooled coefficient that is positive on tbeskill difference term, while tbe absolute valuetnodel estimates an inverse relationsbip be-tween skill differences and our dependent vari-able, real FDI stock. While the coefficient onthe absolute skill difference term is not statisti-cally significant at standard confidence levels,the marginal effect of absolute skill differences(as reported in Table 3) is substantial and sta-tistically significant. At tbe means of the data, astandard deviation increase in the skill differ-ence term (about 2.5 years difference in averageeducational attainment) means a $3.3 billiondecrease in FDI stock by tbe parent country inthe bost country. This is a substantial amount,given a sample average of $4.3 billion FDIstock. These results provide some evidence tbatthe negative relationship between absolute skilldifferences and FDI activity is a worldwidephenomenon, not confined to tbe United States.

III. Conclusions

This paper identifies and corrects an econo-metric specification probletn that led to incor-rect inferences in CMM about the efficacy ofthe knowledge-capital model. The empiricalframework of CMM estimates coefficients ondifference terms that must be interpreted in anopposite manner depending on whether suchdifference terms are negative or positive in value.CMM incorrectly estimate pooled coefficientsover a sample of negative- and posifive-valueddifference terms. Their estimates coincidentlyaffirm the predictions of the knowledge-capitalmodel. However, when a correct specificationof tbe difference terms is employed, either bytaking absolute values or running separate sam-ples for positive- and negative-valued observa-tions, we obtain coefficient signs that do notsupport the knowledge-capital model. Theseresulls aiise not only for the sample of U.S. bilat-eral observations on MNE affiliate sales employedby CMM, but also for U.S. sainples with altema-tive proxies for key variables, as well as a sampleof FDI activity across OECD countries.

Beyond pointing out tbe econometric mis-specification in CMM, this Comment also

shows tbat a variety of databases on MNE ac-tivity show a strong negative relationship be-tween skill dissimilarity and affiliate sales. Thisevidence suggests that Markusen and Venables'(2000) horizontal model of FDI cannot be re-jected in favor of tbe knowledge-capital modelof Markusen et al. (1996). We acknowledge tbatit is possible for skill dissimilarity to be in-versely related to affiliate sales in the knowledge-capital model, but this is true only for certainparameter spaces of the knowledge-capital model.In general, the vertical MNE features of theknowledge-capital model, wbicb would suggestgreater MNE activity for greater skill differences,are at odds witb tbe broad trends in the data. Wecaution that this does not necessarily imply tbatvertical MNE activity does not exist in the realworld, but simply reflects that tbe preponderanceof activity is horizontal in nature or. at least, be-tween the rich countries of the world, whereskill differences are relatively small.''

On a final note, we have used the CMMempirical specification to test the predictions ofthe knowledge-capital model. As CMM note,the underlying theoretical MNE model is quitecomplex, requiring numerical methods to solvea system of over 40 equations. As a matter ofnecessity, one must tberefore work empiricallywith approximations of the model. Tbe problemof identifying the most appropriate empiricalspecification witbin which to test MNE modetimplications merits further research.

APPENDIX

This Appendix provides details on datasources and variable construction for the mod-ified U.S. and OECD data samples used in theeconometric analysis reported in Table 4.

U.S. Modified Sample

As in CMM. we rely on U.S. Bureau ofEconomic Analysis (BEA) data on affiliatesales for our dependent variable. We convertedour affiliate .sales into millions of real U.S.dollars using the U.S. GDP deflator as reported

' For example, see Roben C. Feenslra and Gordon H.Hanson (1999). Mailhew J. Slaughter (2000). and Hanson elal. (2001) for evidence of U.S. verlical MNE acliviiy.

VOL 93 NO. 3 BLONIGEN ETAL: THE MULTINATIONAL ENTERPRISE. COMMENT 991

in the Economic Report of the President and theyearly average exchange rate as reported by theIntemational Monetary Fund's Intemational Ei-nancial Statistics. Affiliate sales are perhaps themost reliable measure of FDI activity (as op-posed to FDI flows or stocks) since they can becompared across time and industries with lessconcern over divergent accounting methodolo-gies. These data are available at the BEA'sInternet site and go back to 1983 for U.S. out-bound sales and 1984 for U.S. inbound sales,though data availability concems with respectto other sample variables limited CMM to usingonly data after 1986. Our alternative proxiesdiscussed next allow our sample to extend backto 1983 and 1984 for U.S. outbound and in-bound affiliate sales, respectively.

For variables using real GDP in both themoditied U.S. and OFCD samples, we use thePenn World Tables real GDP measures avail-able at http://datacentre.chass.utoronto.ca:568Q/pwt/ and described by Robert Summers andAlan Heston (1991). In contrast, the CMM da-tabase constructs real GDP data from the Inter-national Einancial Statistics (IFS) published bythe International Monetary Fund. The PennWorld Tables data only extend until 1992, whilethe IFS data allowed CMM's sample to extenduntil 1994. However, the Penn World Tables goto great lengths to derive real GDP measures inbillions of constant U.S. dollars that ensurecomparability across countries, so it provides aviable alternative to the IFS data.

To construct an alternative proxy for countryskill abundance, we tum to Barro and Lee dataon educational attainment, and define a coun-try's skilled-labor abundance as the average ed-ucational attainment. These data run until 1990.so we repeat 1990 values for 1991 and 1992.Our measure of skilled labor contrasts withCMM's use of annual surveys conducted by theIntemational Labour Organization to constructmeasures of skilled labor to total employmentby country and year. Specifically, their measureis the percentage of total employment that isemployed in categories 0/1 (professional, tech-nical, and kindred workers) and 2 (administra-tive workers). One concem with their data iscomparability of classification schemes acrosscountries (e.g., Japan's share of skilled laborforce averages half that ot the United States andother developed countries). Interestingly, for

the observations in the CMM database the cor-relation between the ILO skill measure and ourBarro and Lee education measure is 0.87.

The variables that provide the largest checkon the CMM sample size because of data avail-ability is the trade and investment cost proxies.CMM rely on data from the World EconomicForum (WEF) wliich provide indicators basedon extensive surveys for a limited number ofyejirs and countries. For our measure of in-vestment barriers, we use the composite scorecompiled by Business Environment Risk Intel-ligence, S.A. (BERI). This composite includesmeasures of political risk, financial risk, andother economic indicators and ranges betweenzero and 100. with higher numbers meaningmore openness. To compare these estimates topreviously used measures of investment barri-ers, we define Investment Barriers as 100 minusthe BERT'S composite score. The BERI measureallows us to consider more countries over alonger time period than CMM. There is a strongrelation between the two, with a correlation of0.81 for the observations in the CMM database.As an alternative to the WEF trade cost mea-sures, we u.se the trade openness measures fromthe Penn World Tables, which are defined as thesum of a country's imports and exports dividedby the country's GDP. We define irade costs as100 minus this trade openness measure. Thecorrelation between this measure of trade open-ness and the WEF for overiapping observationsis much lower than for the investment costproxies, only around 0.20.

The resulting U.S. moditied sample, usingthese altemative proxies for real GDP, skilled-labor abundance, and trade and investmentcosts, covers 51 countries and years from 1983-1992. Appendix Table Al gives summary sta-tistics for the variables used in the U.S.modified sample.

OECD Sample-

Unfortunately, very few OECD countrieskeep track of affiliate sales, and there is nocomprehensive cross-country database of for-eign affiliate sales activity, even for OECD coun-tries. Thus, when considering the OECD siunple,we must resort to data on bilateral FDI stocks asreported by OECD-member countries. These dataare reported in the OECD's Intemational Direct

992 THE AMERICAN ECONOMIC REVIEW JUNE 2003

TABLF. A 1 -—SUMMARY

Variables Observations

Real Affiliate SalesGDP SumGDP Difference SquaredBDH Skill Difference

BDH Skill"Difference * GDPDifference

Absolute BDH Skill DifferenceAbsolute BDH Skill Difference

* Absolute GDP DifferenceBDH Investment CosI HostBDH Trade Cost HostBDH Trade Cost Host *

Squared Skill DifferenceBDH Trade Cost ParentDisiance

778778778778

778

778778

778778778

778778

STATISTICS FOR

Mean

20,8214,523.0

l.59e+07-0.04

18.349

4.6718.579

37.858.7

1.723.6

56.05,077.8

MODIFIED U.S. SAMPLE

Standard deviation

40,630484.2

3,377,8505.19

9,687.6

2.239,236.9

13.344.2

1,924.8

46.12,370.8

Minimum

03.610.2

5,277,082-9.51

-6,407.9

0.32744.3

17.3-278.8

-5,335.8

-278.8455.0

Maximum

267,4016,449.0

2.09e+079.51

41.584

9.5141,584

70.491.0

7,337.4

91.010.163.0

Notes: p = parent, h = best. Affiliate sales and FDI stock measured in millions of real U.S. dollars. GDP terms in billionsof real U.S. dollars.

TABLE A2—COVERAGE OF OECD FDI STOCK DATA

Country

Australia

Austria

Canada

FinlandFrance

Italy

Japan

Netberlands

Norway

United Kingdom

United States

Direction of FDI stocks

tnboundOutboundInboundOutboundInboundOutboundInboundInboundOutboundInboundOutboundInboundOutboundInboundOutboundInboundOutboundInboundOutboundInboundOutbound

Number of partnercountries reported

17129

1129301

43342223

371414IS2419352841

Years covered

1982-1982-

1982. 1986-1982, 1986-

1982-1982-1989-1987-1987-1985-1985-1982-1982-1984-1984-1987-1988-1982-

1984, 1987-1982-1982-

Notes: Number of partner countries reported are for 1990. Not all reported countries arenecessarily reported each year during range indicated.

Investment Statistics Yearbook."^ Since the dataare collected from national sources in each coun-

'" These data are available in print form in these annuaiyearbooks or in electronic form on (he OECD StatisticalCompendium CD-ROM, available for purchase from theOECD.

try, there is substantial variation in coverage bycountry source and by year, and there is variationin measurement of FDi activity itself. On thewhole, about half of the OECD countries reportmeasures of inward and outward stocks of FDI forsome countries and for some years. The earliestdata available begin in 1982. Appendix Table A2

VOL 93 NO. 3 BLONIGEN ET AL: THE MULTINATIONAL ENTERPRISE, COMMENT 993

TABLB A3—SUMMARY S'I .ATISTICS FOR OECD

Variables

Real FDI StockGDP SumGDP Difference SquaredBDH Skill Difference

( = EduCp - Educ,,)BDH Skill Difference * GDP

DifferenceAbsolute BDH Skill DifferenceAbsolute BDH Skill Difference

* Absolute GDP DifferenceBDH Investment Cost HostBDH Trade Cost HostBDH Trade Cost Host *

Squared Skill DifferenceBDH Trade Cost ParentDistance

Observations

2.4602.4602.4602,460

2,460

2.4602,460

2.4602,4602,460

2,4602,460

Meati

4.321.51.674.1

3.156,3241.65

3,401.2

2.553,850.4

42.03t.3

422.6

52.06,302.6

Standard deviation

11.7621.497.7

5,820,7882.69

6.459.6

1.866,202.3

12.359.3

1,050.1

22.34,791.6

Minimum

-357.173.0

0-5.40

-6.995.7

O.OI0.03

17.3-286.2

-6.559.1

-18.8174.0

Maximum

176,7816.449.0

2.09e+()78.10

31.() I!

8.103t,0ll

65.087.3

5.599.5

82.0418.372

Notes: p = parent, li = host. Affiliates sales and FDI stock measured in millions of real U.S. dollars. GDP terms in billionsof real U.S. dollars.

provides further details on data coverage acrossOECD countries and years in our sample." Weconverted our FDI variables into thousands of realU.S. dollars using the U.S. GDP deflator as re-ported in the Eeonomic Report of the Presidentand the yearly average exchange rate as reportedby the International Monetary Fund's Interna-tional Financial Statistics.

For control regressors. we use the Blonigen-Davies-Head (BDH) altemative proxies de-scribed above for the U.S. modified sample.Combining this with the OECD data on FDIoutbound stock across countries and time v 'ehave a database that spans 15 OECD parentcountries, and 38 host countries (some OECD

" Tfiere are some comparability concerns of FDI mea-sures acro.ss countries. For example, a number of OECDcountries do not include reinvested eamings by firms intheir measures of FDI. Countries can also differ in whatpercentage of foreign-owned shares of a firm arc necessaryfor it to be classified as FDI raiher than portfolio investnienl.IMF and OECD guidelines specify investment as FDI whenacquired shares are 10 percent or higher of target firm'soutstanding stock, which many of the countries foUow oreventually adopted. Edward M. Graham and Paul R. Krug-man (1995) tind that the foreign parent of a MNE in lheUnited States on average owns 77.5 percent of the affiliatesequity, suggesting that this problem may not be overwhelm-ing. However, with only a couple of exceptions, we notethat FDI definitions are fairly consistent for the same coun-try over time.

and some non-OECD) over the years 1982through 1992.'" This leaves 2,460 observationsand Appendix Table A3 provides summary sta-tistics for the variables used in the OECDsample.

REFERENCES

Brainard, S. Lael. "An Empirical Assessment ofthe Proximity-Concentration Trade off Be-tween Multinational Sales and Trade." Amer-ican Economic Review, September 1997,H7(4), pp. 520-44.

Carr, David L.; Markusen, James R. and Maskus,Keith E. "Estimating the Knowledge-CapitalModel of the Multinational Enterprise."American Economic Review, June 2001,97(3), pp. 693-708.

Ekholm, Karolina. Multinational production andtrade in technological knowledge. Lund Eco-nomics Studies (Lund. Sweden) No. 58.Lund: University of Lund, March 1995.

'- We gathered data on outbound FDI stock only, sincemost of lhe reported data on OECD FDI activity is betweenOECD countries, with some information on FD! stock ofOECD countries in non-OECD countries that involve sub-stantial FDI activity. Inbound FDI stock daia woiik! oniyrevea! new observations of non-OECD investment intoOECD countries, for which there were relatively few in-stances of such recorded data.

994 THE AMERICAN ECONOMIC REVIEW JUNE 2003

. "Factor Endowments and the Pattemof Affiliate Production by Multinational En-terprises." CREDIT (University of Notting-ham) Working Paper No. 97/19, 1997.

. "Headquarter Services and RevealedFactor Abundance." Review of InternationalEconomics, November 1998a, 6(4), pp. 545-53.

. "Proximity Advantages, Scale Econo-mies, and the Location of Production," inPontus Braunerhjeim and Karolina Ekholm,eds., The geography of multinational firms.Boston, MA: Kluwer Academic Publishers,1998b, pp. 59-76.

Feenstra, Robert C. and Hanson, Gordon H. "TheImpact of Outsourcing and High-TechnologyCapital on Wages: Estimates for the UnitedStates, 1979-1990." Quarterly Joumal ofEconomics, August 1999, 114(3). pp. 907-40.

Frankel. JeWrey. "On the Mark: A Theory ofFloating Exchange Rates Based on Real Tn-terest Differential." American Economic Re-view, Septetnber 1979, 69(4), pp. 610-22.

Graham, Edward M. and Krugman, Paul R. Eor-eign direct investment in the United States,third edition. Washington, DC: Institute forIntemational Economics, 1995.

Hanson, Gordon H.; Mataloni Jr., Raymond J.and Slaughter, Matthew J. "Expansion Strate-gies of U.S. Multinational Firms." NationalBureau of Economic Research (Cambridge,MA) Working Paper No. 8433, August 2001.

Haynes, Stephen E. and Stone, Joe A. "On theMark: Comment." American Economic Re-view, December 1981, 7/(5), pp. 1060-67.

Heipman, Elhanan. "A Simple Theory of Inter-national Trade vi/ith Multinational Corpora-tions." .lournal of Political Economy, June1984, 94(3). pp. 451-71.

Horstmann, Ignatius J. and Markusen, James R."Strategic Investments and the Developmentof Multinationals." International EconomicReview, February 1987, 28(\), pp. 109-21.

"Endogenous Market Structures in In-ternational Trade." Journal of International

Economics, February 1992, 52(1/2), pp. 109-29.

Markusen, James R. "Multinationals, Multi-plant Economies, and the Gains from Trade."Journal of International Economics, May1984. /6(3/4), pp. 205-26.

"Trade versus Investment Liberaliza-tion." National Bureau of Economic Re-search (Cambridge, MA) Working Paper No.6231, October 1997.

Markusen, James R. and Maskus, Keith E. "Dis-criminating Among Altemative Theories ofthe Multinational Enterprise." National Bu-reau of Economic Research (Cambridge,MA) Working Paper No. 7164, June 1999.

-. "Multinational Fimis: ReconcilingTheory and Evidence," in Magnus Blom-strom and Linda S. Goldberg, eds.. Topics inempirical intemational economics: Afest.schrift in honor of Robert E. Lipsey. Chi-cago: University of Chicago Press, 2001, pp.71-95.

Markusen, James R. and Venables, Anthony J,"The Role of Multinational Firms in theWage-Gap Debate." Review of InternationalEconomics, November 1997, 5(4), pp. 435-51.

. "Multinational Firms and the NewTrade Theory." Joumal of International Eco-nomics, December 1998, 46(2), pp. 183-203.

. "The Theory of Endowment, Intra-Tndustry and Multinational Trade." Journalof Intemational Economics. December 2000,52(2), pp. 209-34.

Markusen, James R.; Venabies, Antbony J.; Eby-Konan, Denise and Zbang, Kevin Honglin. "AUnified Treatment of Horizontal Direct In-vestment, Vertical Direct Investment, and thePattern of Trade in Goods and Services."National Bureau of Economic Research(Cambridge, MA) Working Paper No. 5696,August 1996.

Slaugbter, Mattbew J. "Production Transferwithin Multinational Enterprises and Ameri-can Wages." Joumal of International Eco-nomics, April 2000, 50(2), pp. 449-72.