Knowledge Seeking and Location Choice of Foreign ... - CiteSeerX

Knowledge Seeking and Location Choice of Foreign Direct Investment in the United States Wilbur Chung • Juan Alcácer

The Wharton School, University of Pennsylvania, 2027 Steinberg-Dietrich Hall, Philadelphia, Pennsylvania 19104 Stern School of Business, New York University, New York, New York 10012 [email protected] • [email protected]

T

o what extent do firms go abroad to access technology available in other locations? This paper examines whether and when state technical capabilities attract foreign investment in manufacturing from 1987-1993. We find that on average state R&D intensity does not attract foreign direct investment. Most investing firms are in lower-tech industries and locate in low R&D intensity states, suggesting little interest in state technical capabilities. In contrast, we find that firms in research-intensive industries are more likely to locate in states with high R&D intensity. Foreign firms in the pharmaceutical industry value state R&D intensity the most, at a level twice that of firms in the semiconductor industry, and four times that of electronics firms. Interestingly, not only firms from technically lagging nations, but also some firms from technically leading nations are attracted to R&D intensive states. This suggests that beyond catching up, firms use knowledge-seeking investments also to source technical diversity. (FDI; Location Choice; Knowledge Seeking; Random Parameter Logit)

Introduction

Firms have different motives for expanding abroad including sourcing low-cost factors, avoiding taxes, managerial rent seeking, and behaving strategically by following competitors.1 While numerous, one motive has been prominent in prior research: Firms conduct foreign direct investment (FDI) when they possess unique capabilities. Although firms might sometimes prefer to sell their unique capabilities to others, market failures prevent their transfer to foreign firms. As a consequence, firms with unique capabilities expand abroad themselves—their unique capabilities are “internalized” to obtain higher returns. Recently, another prominent motive has emerged. Instead of utilizing capabilities already on 1

For a review see Caves (1996).

Management Science © 2002 INFORMS Vol. 48, No. 12, December 2002 pp. 1534–1554

hand, firms may expand abroad in search of capabilities that are not available in their home markets (Cantwell 1989). This motive has been termed “technology seeking” or “knowledge seeking.” Most empirical evidence of knowledge seeking as a motive for FDI comes from studies that explore location of R&D facilities in research-intensive industries. While these studies clearly demonstrate the existence of knowledge-seeking behavior, we still know little on the relative importance of knowledge seeking versus internalization as a motive for FDI across industries. How prevalent is knowledge seeking? Is knowledge seeking a phenomenon particular to specific research-intensive industries or is it common across all industries? Can knowledge seeking induce FDI in manufacturing? An additional question is central to understand knowledge seeking: What types of firms are more 0025-1909/02/4812/1534$5.00 1526-5501 electronic ISSN

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

likely to invest abroad to acquire technology? The conventional wisdom is that knowledge seeking occurs among technical laggards trying to reduce their gap by investing abroad to acquire needed knowledge. Recently, Cantwell and Janne (1999) supplement this prevailing view by suggesting that firms from leading technical centers will go to other leading technical centers not to catch up but to increase their knowledge diversity. Understanding knowledge seeking is critical for managers and policy makers alike. Managers should enhance existing safeguards that protect proprietary knowledge if foreign firms’ motive is acquiring local technology. For example, they might reconsider forming certain alliances or they might provide stronger incentives to retain key employees. For policy makers, the presumption has been that inward FDI is beneficial—that FDI increases competition and productivity. However, if many foreign firms enter seeking new knowledge, these gains may not accrue and a nation’s technological uniqueness might be more quickly replicated. To answer these questions, we acknowledge that firms have different motives for investing abroad and, as a result, they will value the knowledge available in a location differently. To account for this heterogeneity in valuation, we focus on the observable location choices of inward FDI: Those investments by firms seeking knowledge will be drawn to locations with distinctly different factors than firms following other motives. We expect firms seeking knowledge to value locations that offer more technical activity, where there are more scientists, more patents being generated, and where there is greater R&D intensity. Firm traits will influence the likelihood of seeking knowledge; e.g., knowledge seeking may be more prevalent if a firm is in an R&D intensive industry like pharmaceuticals or electronics. Additionally, following the classic knowledge-seeking view, the more a firm’s technical capabilities lag behind, the more likely the firm will be to seek technology. In contrast, firms pursuing other motives will be disinterested in the technical activity in a location; instead these firms will seek low factor costs, better market access, greater market demand, or other traits. Management Science/Vol. 48, No. 12, December 2002

Our statistical analysis is based on the random parameter logit (RPL), a generalized version of the conditional logit. Unlike the conditional logit, which assumes that firms value location traits equally, the RPL allows preferences to vary across firms by assuming a random distribution of coefficients. With this technique we are able to determine in which industries knowledge seeking occurs, and to quantify differences in firm valuations of location traits. Our empirical context is manufacturing FDI from the Organization for Economic Cooperation and Development (OECD) nations into the United States for 1987–1993. We examine in which of the 48 contiguous states 1,784 inward FDI transactions locate. We use a state-level analysis because of data constraints and because political boundaries have implications for economic activity (tax rates, labor laws, and other traits are set at the state level). Recognizing that states are a relatively broad geographic classification and that economic activity may not follow these politically defined boundaries, we also conduct a secondary analysis using finer geographic gradations. This analysis, based upon 170 economic areas defined by the Bureau of Economic Analysis, provides findings consistent with the state-level analysis. We find that certain technical activity uniformly attracts inward FDI, notably the count of doctorates scientists and engineers in a state. Surprisingly, state R&D intensity and the count of patents awarded to state residents decrease inward FDI, suggesting that knowledge seeking is not prevalent across industries. Importantly, the variation of the negative effect of state R&D intensity on inward FDI is significant and large enough to indicate that while most firms are not attracted by state R&D, some are. Firm traits explain most of this variation. We find that knowledgeseeking activity is limited to only R&D intensive industries. Foreign firms investing in pharmaceuticals, semiconductors, and electronics have positive valuations, while firms investing in chemicals have negative valuations of state R&D intensity. Interestingly, knowledge seeking occurs not only among firms trying to catch up: In pharmaceuticals we find that it is not the technical laggards but the firms from leading technical nations that more highly value state R&D intensity. 1535

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

We proceed as follows. The next section discusses knowledge seeking and develops in which settings firms are more likely to pursue this motive. Then we discuss the random parameter logit with an emphasis of its advantages over other methods and how the features of the RPL provide greater richness to test hypotheses. Next are the data and results. Finally, we highlight several findings, discuss future research, and conclude.

The most recognized reason for firms to conduct FDI is that they possess unique capabilities that can be deployed abroad. Although firms might prefer to sell their unique capabilities, market failures prevent firms from transferring them to others. Therefore, firms obtain the highest value by expanding abroad themselves and keeping these useful capabilities internal to the firm.2 Consistent with the importance of “internalization,” Morck and Yeung (1991) show that stock markets react positively to international expansion by firms that possess larger stocks of intangible assets such as technologies developed through R&D spending. More recent research suggests another motivation for FDI—seeking technology. Instead of utilizing capabilities already on hand, firms may expand abroad in search of knowledge and skills. Cantwell (1989, pp. 138–139) notes that technology differs across locations because technology depends on location–specific factors, such as innovations previously established, the education system, and the linkages between educational institutions and firms. As a consequence, firms may supplement their existing technologies by expanding internationally to access new knowledge (Cantwell 1989, Wesson 1993). Accessing localized knowledge requires physical proximity because, as Kogut and Zander (1992) argue, some knowledge is partially tacit and transfer requires frequent interaction. For example, Jaffe et al. (1993) use patent citations to show that inventors

are more likely to cite other inventors who are geographically proximate. Extended to an international context, Almeida (1996) shows that foreign firms in the semiconductor industry cite same-region patents more often than local firms in a comparable control group, which suggests that foreign firms make greater use of local knowledge than local counterparts. Shan and Song (1997) show that in biotechnology, foreign firms make equity investments in United States biotechnology firms with larger patent stocks to draw upon local knowledge. Most empirical evidence of knowledge seeking comes from the internationalization of R&D activities in research-intensive industries.3 Kuemmerle (1999) argues that firms in biotechnology establish R&D facilities to both “exploit” and “augment” their R&D capabilities. Florida (1997) finds that accessing new indigenous technology is more important than customizing existing technology for new markets in a sample of 207 foreign research laboratories in the United States in biotechnology/drugs, electronics, chemicals/materials, and automotive. For a similar set of industries, Serapio and Dalton (1999) integrate several data sources on foreign research labs in the United States and argue that firms sometimes invest to access new technology. Firms may also seek technology in less researchintensive industries and not only through R&D facilities, but also through manufacturing operations. Cantwell (1989, p. 8) argues that: “The acquisition of new skills, and the generation of new technological capacity, partially embodied in new plant and equipment, must be a goal of every firm


.” Empirical evidence for this claim is more scattered. Kogut and Chang (1991) look across manufacturing industries in the United States and show that more Japanese FDI transactions occur in industries that have greater R&D differences, a finding that is consistent with “sourcing of US locational advantages in technology.” Similarly, Pugel et al. (1996) show that the share of industry value produced by foreign firms increases with host country R&D and not with home country R&D, which is consistent with the knowledge-seeking argument.

2

3

Location Choice with Heterogeneous Motives

Internalization theory is developed by Coase (1937), Caves (1971), and Buckley and Casson (1976), among others.

1536

Research Policy (1999 28(2, 3)) reflects recent advances in the growing literature in internationalization of R&D activities.

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Underlying these studies is the assumption that knowledge seeking is only for firms that want to catch up. Cantwell and Janne (1999) challenge this assumption by suggesting two types of knowledge-seeking behavior. They differentiate between firms originating from leading versus lagging technical centers. Firms from lagging technical locations do need to catch up and locate labs abroad with an emphasis on improving their existing technology. Although firms from leading locations do not need to catch up, they still may locate labs abroad to source more diverse knowledge. This suggests that knowledge seeking might also occur in situations where differences in R&D between nations are small. Leading technical centers will have little difference in R&D intensities, yet substantial cross-investment may occur to source technical diversity. Although previous research begins to develop our understanding of heterogeneous motives for FDI, we still know little about their relative importance across industries. Is knowledge seeking as important for non-research-intensive industries as it is for researchintensive ones? Can knowledge seeking motivate FDI in manufacturing? Are all knowledge-seeking firms technical laggards trying to catch up, or is knowledge seeking a behavior present in all firms in certain industries? To address these questions we need to first acknowledge that firms have different reasons for expanding abroad. When following more traditional motives, entrants will be interested in maximizing profit, e.g., maximizing revenues and minimizing costs. For example, firms that are internalizing will want to use their capabilities abroad as effectively as possible and will be attracted not only to locations that offer low-cost operations, but also greater revenues. In contrast, knowledge-seeking firms will locate close to sources of knowledge. For example, Almeida and Kogut (1999) use patent citations to show that the mobility of engineers strongly determines how localized knowledge is. They argue that because much knowledge is tacit, that knowledge resides with engineers in local labor markets. Therefore, firms seeking new knowledge will have to be able to access such local markets and locations with Management Science/Vol. 48, No. 12, December 2002

greater technical activity will be more attractive to firms seeking knowledge. We assume that firms choose the location that maximizes their utility. Because firms can expand abroad for numerous reasons, this utility is likely composed of several categories. Importantly, firms are unlikely to equally value all categories. Clearly, the presence of technical activity, market size, market access, and lowcost factors are not going to be equally attractive. To account for this heterogeneity in the valuation of location attributes, we introduce firm traits. Firm traits will strongly affect the value placed on categories in the utility function. An important firm trait will be the technical capabilities of its home country—whether the home industry is leading or lagging technically. Because knowledge is localized, the technical capabilities in their home country-industry will be a strong determinant of indigenous firms’ technical capabilities; firms are unlikely to deviate far from the country knowledge base. Therefore, a firm’s country-industry technical capabilities can help identify what motive the firm is more likely to follow. To pursue internalization, a firm needs some unique capabilities that are difficult to transfer through the market. While a firm originating from a country-industry with leading technical knowledge may have such unique capabilities, a firm originating from a country-industry with lagging technical knowledge is less likely to be so endowed. This is not to say that no firm from less technically capable country-industries will ever have unique capabilities; rather such firms will be exceptions rather than the norm. Conversely, firms from country-industries that are technical laggards are more likely to seek knowledge abroad. Both Kogut and Chang (1991) and Kuemmerle (1999) find that R&D difference at a country level is a determinant of knowledge seeking. Again, this is not to say that firms from leading technical country-industries will never knowledge seek, but on average firms from lagging country-industries are more likely to do so. A second important trait will be the overall knowledge intensity of an industry. While a firm might lag technologically, knowledge seeking is less necessary if the firm is in an industry where knowledge composes a smaller portion of the value added. Firms in 1537

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

mature industries in which standard and well-known technology is used may seek to compete on quality, service, or other less technical dimensions. In contrast, there may be other industries where knowledge is crucial. Advancing technical knowledge may be the basis of competition. Firms constantly strive to outpace each other through innovation. In such settings, not only laggards, but all participants will need to be aware of competitors’ technical activities. Monitoring at arms’ length may be inadequate, leading many industry participants to collocate. Similarly, economies of scale in development of technology might be so great that firms need to obtain knowledge wherever they can. Cantwell and Janne (1999) suggest that firms from leading technical centers will go to other leading technical centers not to catch up, but to increase their knowledge diversity. Thus in certain industries, firms that are technical leaders will expand abroad to seek knowledge. In summary, we expect some entrants to value locations’ traits reflecting localized technical activity and others to value traits reflecting market demand, lowcost factors, and market access. The value that firms place on some traits may vary greatly; some traits may be highly valued by some but valueless to others. Importantly, entrants’ own traits should account for some of this variation. Specifically, we expect firms from country-industries with relatively greater technical capabilities will be more attracted to locations with greater market demand and market access, while firms from country-industries with relatively lower technical capabilities will be attracted to locations where more technical activity occurs. Finally, we expect that firms in certain industries—where technical progress is critical—will be attracted to locations of greater technical activity, regardless of their technical capabilities.

Method

Our hypotheses encompass how attractive location traits are to entering firms, how this attractiveness varies by firm, and how this variation is affected by firm traits. Given these expectations we can specify a general utility function for firms choosing from 1538

among a set of locations. The utility of firm i from location j is: uij = xj ij + ij 

(1)

where xj is a row vector of location j characteristics such as technical activity and market demand, ij is a column vector of firm-specific taste coefficients for each location, and ij is a mean zero stochastic term that is distributed i.i.d extreme value. With this basic utility function, the influence of firm-specific characteristics enters through ij . Essentially, ij varies by firm and by location, with firm-specific characteristics accounting for part of this variation. We can think of ij as having an average value, which for location j is constant for all firms, j , and then a firm-specific deviation ij . In turn, this firm-specific deviation results from firm-specific characteristics expressed as a row vector, wi , which are multiplied times a column vector of constants, j , which are called “deep parameters,” ij = j + ij

where ij = j wi


(2)

Because Equations (1) and (2) specify that ij is different for every firm, we need to impose some structure for estimation. We can assume that in aggregate, such firm-specific taste coefficients can be represented as an overall distribution with a mean and a standard deviation. Typically, this distribution is either assumed normal or lognormal; e.g. ij is ∼ normal or log-normal [j + ij , 2 ].4 Then the statistical task is to estimate the parameters of ij —the mean (j + ij , the standard deviation ( 2 , and deep parameters (j for each element of the column vector. In this case, the deep parameters correspond to firm traits that explain variation in valuation. To test this utility function, we use the random parameter logistic regression (RPL). The RPL is a generalization of the traditional conditional logit that has been used recently by Berry et al. (1995), Revelt and Train (1998), and Villas-Boas and Winer (1999). The probability that firm i chooses location j from the set 4

A normal distribution allows estimation of means for ij that are either positive or negative, while a lognormal distribution constrains the means to be only either positive or negative.

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

of locations k, conditional on i (each firms’ specific attributes) is Prchoice j  i = Lij i =


ei xij
i xik k=j e

(3)

The unconditional probability is then the conditional probability under all possible conditions. Assuming i is distributed with density f  , then multiplying the conditional probability by this density and integrating gives each firms’ unconditional probability density:  (4) Pi  = Lij i f  di
Using Equation (4), the log-likelihood function is    Log L = ln Pi  = ln Lij i f  di
(5) i

i

Exact maximum likelihood estimation (MLE) of Equation (5) is not possible because the integral cannot be calculated analytically, which leads to numeric approaches. Similar to Revelt and Train (1998), we approximate the probability, Pi  , through simulation and maximize the simulated log-likelihood function. Greene (2000, pp. 872–873) shows that this probability can be approximated using the fact that the expected value of each firms’ unconditional probability density can be estimated numerically with repeated draws: R 1 Pr(choicej  i eir

 R→ R r=1

EPi   = lim

(6)

where i eir is an estimate of the true i and eir is the error for the rth draw out of R. For each firm there are R draws (typically 100 or greater). This provides an unbiased estimator of Pi  , whose variance decreases as the repeated draws, R, increases. Then the log-likelihood function has NR draws where N is the number of investment transactions, and is given by   N R   1 ln Pi  = ln P  Log L = R r=1 i i i=1   N R  1 = ln Pr(choicej 
i eir
(7) R r=1 i=1 Management Science/Vol. 48, No. 12, December 2002

Equation (7) is the function that then is maximized by traditional MLE. This is the unconditional probability of observing the particular set of firms’ location choices—which firms chose which locations. The RPL requires greater computing power because of the many draws. To have confidence in the estimation results, we follow Revelt and Train (1998) and have 100 draws (R is equal to 100) for each N . While computing power is the main cost, the RPL offers researchers an important gain: The researcher can observe whether independent variables uniformly affect the dependent variable—in our case, whether and which location traits firms value uniformly versus differentially. The independent variable is uniformly important when the mean of the parameter estimate is significant but the standard deviation is not. When the standard deviation is significant, then the independent variable is not uniformly important. Obtaining parameter estimates for both the mean and standard deviation of the locations’ characteristics provides a greater richness for the phenomena being explored. This advantage becomes more evident in our subsequent empirical test.5

Data

We examine the state location choice of 1,784 FDI transactions entering the United States between 1987– 1993 from OECD nations whose primary function is manufacturing.6 The data for FDI transactions comes from the International Trade Administration (ITA) report of “Foreign Direct Investment in the United 5 Alternately, the investigator could use conditional logit models (CLM) that do not allow variation in valuation by actor and therefore would not require simulated maximum likelihood. The RPL results we report subsequently have analogous CLM results that are very similar. The main gain from using RPL is accounted for, revealed, and explained unobserved heterogeneity. 6 We do not use any FDI transaction data from before 1987 due to concerns with comparability of SIC codes—1987 was a major revision year for United States SIC codes. To maintain comparability, we only use 1987 forward. This FDI transaction data is also bounded above; the ITA stopped collecting the data in 1994. We restrict ourselves to OECD nations because data on firm R&D spending and intensity is available. Inward FDI from non-OECD nations in manufacturing account for only 18 additional transactions during this period.

1539

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

States, (various years) Annual Transactions.” While by no means exhaustive, the ITA uses a wide range of public sources to assemble the annual list. This is the same data source used by Kogut and Chang (1991), Hennart and Park (1993), Blonigen (1997), Shaver (1998), and others to construct their dependent variables or focal independent variables. For each investment transaction, the ITA lists the mode of entry, transaction value, country of origin, fourdigit SIC (Standard Industrial Classification) industry/function, location by city and state, and foreign owner. We use only those transactions whose primary function is manufacturing, those with SICs between 2000–3999.7 We use both acquisitions and greenfield manufacturing investments, assuming that mode of entry does not constrain location choice; e.g., potential acquisition targets will exist in most desired locations and firms can build new greenfield facilities.8 Our dependent variable is the location for each inward FDI transaction. Location can be defined broadly, like New England or the Pacific Northwest, or more narrowly, like the Bay Area or Detroit, MI. Our selection of the level of analysis is driven by several concerns. First, the level chosen must be suitable for testing our theoretical relationships. Knowledge seeking is based upon knowledge being partially tacit and requiring frequent contact for transfer; suggesting physical proximity and a finer-grained level of analysis. Alternately, if firms are primarily concerned with minimizing costs, then some measures such as 7

Depending upon how much information was available for the ITA, the four digit-SIC indicates the investing firm’s industry or the transaction’s purpose. For example, while Mazda Motor Manufacturing of Japan (primary SIC 3711) made five investments into the United States in 1987, the ITA was able to identify that two were for sales subsidiaries (which they assigned to SIC 5012) and one was a research lab (which they assigned to SIC 7391). While separating out sales subsidiaries and R&D labs, the ITA does not consistently do so; overall, very few research labs have been declared, which is consistent with the ITA classifying an investment as manufacturing when both manufacturing and R&D are present. Given the noisy nature of this industry/function measure, the RPL is especially suitable because it can address heterogeneity resulting from industry/function without having industry/function clearly indicated. 8

Results with and without investment by acquisitions are similar. We discuss this further in the robustness section.

1540

tax rates and right-to-work laws that are defined at the state level may be important. A final concern is data availability. We need numerous explanatory variables to reflect a full spectrum of locations’ attributes that will draw FDI following heterogeneous motives. Given the paucity of data at other than the state level, we use the state as our primary level of analysis but also conduct a secondary analysis at a slightly finer geographic level. Using state location as our dependent variable, our choice set is composed of 45 states and the District of Columbia because we include only the 48 contiguous states, and three states had no inward investments during our investigation period (Idaho, North and South Dakota). The location of these 1,784 transactions is shown in as Figure 1. We also use a finer geographic unit, the Economic Area (EA), to complement the location analysis by state. The Bureau of Economic Analysis (BEA) uses counties as building blocks to define 170 economic regions that span the continental United States.9 Counties are grouped based on commuting patterns of employees to maximize the percent of people that live and work within the same region. Figure 2 shows the transactions in our sample at the economic area level. Unsurprisingly, many investments, 32% of the sample, fall into four major metropolitan areas: NYC (EA10), San Francisco (EA163), LA (EA160), and Chicago (EA64). For state characteristics, we start with the set of variables used by Coughlin et al. (1991) that captures three dimensions for each state: market size, access to surrounding markets, and cost of production factors. Population density, income per capita, and tax per capita measure market demand. Access to surrounding markets is captured by two variables: the number of airports per capita and miles of highways per capita. High values for these two measures indicate low connectivity—a state will have a 9

Identifying Economic Area (EA) for an investment transaction requires two stages of matching. First, the transaction’s city and state need to be matched to the appropriate county. Second, the county is then associated with an EA. Due to difficulties in both stages, we were able to match only 1,660 of the 1,784 transactions to EAs. Surprisingly, the concordance of county-to-EA provided by the BEA is not exhaustive; numerous counties are missing.

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Figure 1

Location of Inward FDI Transactions 1987–1993 by State

Note. There are 1,784 transactions from OECD nations in manufacturing industries to the 48 contiguous states and the District of Columbia shown.

high number of highway miles per capita because it has to have a minimal level of highway infrastructure for a small population—such states are relatively isolated. Seven measures reflect cost of production factors for each state: land available,10 percent of population employed in manufacturing, unemployment rate, average weekly wage, tax as percent of income, presence of right-to-work laws, and percent of unionized workers.11 All these measures vary by state and by year with the exception of percent of unionized workers. For this indicator, we impute data from the closest year available to missing data points. The 10

The state size in square miles of land less federally controlled land.

11

We make some slight modifications to Coughlin et al.’s (1991) variable list by adding population density and scaling several variables by population instead of by land area. Otherwise, these variables proxy for population density. For example, while Coughlin et al. use manufacturing employees/land area, we use manufacturing employees/population; a state will have high manufacturing employees/land area simply because the state has a large population.

Management Science/Vol. 48, No. 12, December 2002

Appendix lists in detail these variables, their sources, and years of coverage. Given our interest in knowledge seeking, we supplement these data with measures of state technical activity from the National Science Foundation (NSF) report on “Science and Engineering State Profiles.” We use the NSF report for 1997, which details state characteristics mostly from 1995.12 We use two measures to indicate state technical stock: number of doctorates in science and engineering, and R&D intensity (total R&D spending by government, industry, and academia scaled by state gross product). Although a state may have a large stock of technical activity, this stock may lie idle without generating new knowledge. Therefore, we also include the number of science and engineering doctorates earned per year and the number of patents issued to state residents per year to reflect pace of technical creation. 12

Because no other time-varying data is available for earlier years, we assume that these characteristics do not change substantially across time when comparing across states.

1541

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Figure 2

Location of Inward FDI Transactions 1987–1993 by Economic Area

Note. Prepared by Regional Economic Analysis Division, U.S. Department of Commerce, Economics and Statistics Administration, Bureau of Economic Analysis. There are 1,784 transactions from OECD nations in manufacturing industries to 170 Economic Areas (EAs) in the continental U.S. (less 124 transactions where no EA is identified). Economic Area identifier number is in bold, followed by number of transactions into that EA. Map available in color from www.bea.doc.gov/bea/regional/docs/econlist.htm . 170 167 166 164 163 162 161 160 158 157 156 154 153 152

1542

26 33 2 6 129 3 31 119 11 2 2 1 2 4

151 148 147 141 139 138 134 132 131 130 127 126 125 124

5 1 3 21 1 1 7 5 38 3 39 1 4 3

122 121 118 113 109 108 107 106 104 102 101 100 99 98

4 1 2 1 1 1 20 2 1 1 2 6 12 1

97 96 95 90 86 85 84 83 82 80 79 78 77 75

1 11 4 6 1 1 6 4 1 4 2 7 2 1

74 73 72 71 70 69 68 67 66 65 64 63 62 61

5 9 1 25 11 4 3 38 4 6 86 5 12 1

60 59 57 56 55 54 53 52 51 50 49 48 47 46

2 5 58 5 35 2 28 5 21 11 25 6 14 1

45 44 43 42 41 40 39 37 36 34 33 31 30 29

3 11 11 1 9 49 1 1 1 13 1 7 2 1

27 26 25 24 23 21 20 19 18 17 15 14 13 12

1 1 3 4 42 1 10 15 20 2 18 7 29 49

11 10 9 8 7 6 5 4 3 2

5 199 3 9 11 5 7 3 66 1

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Table 1

Descriptive Statistics of State and Firm Characteristics

State characteristic∗

State’s LAND area Population DENSITY Per capita INCOME TAX per capita AIRPORTS per capita HIGHWAY miles per capita % of pop employed in MANUF % UNEMPLOYMENT Average weekly WAGE Has right-to-work laws % UNIONS TAX as % income Count of doctorates in Sci. & Eng. Doctorates EARNED in Sci. & Eng. State R&D INTENSITY (%) PATENTS awarded residents

sq. miles (000s) pop (000s)/sq. miles 000s of dollars percentage count/pop (000s) count/pop (000s) fraction percentage 000s of dollars/week 0 or 1 dummy percentage percentage 000s 000s percentage count

n

Mean

Std. dev.

Minimum

Maximum

min

max

414 414 414 414 414 414 414 414 414 414 414 414 414 414 414 414

3016414 7.57E-04 1802 052 008 002 007 618 044 037 1946 283 1155 058 2381 1315

2563013 3.09E-03 376 027 005 002 003 172 006 048 1180 116 1255 066 1708 1807

2800 1.19E-05 990 008 001 000 002 240 030 000 290 052 084 004 418 33

16597300 2.28E-02 3121 133 029 009 013 1310 072 100 5360 516 7183 339 8717 10473

DC WY MS TN TN NJ WY NE MS

TX DC DC DC WY WY NC LA MI

SC NH WY ME LA WY

MI NY CA CA NM CA

n

Mean

Std. dev.

Minimum

Maximum

FDI transaction characteristics (from 1987–1993)

Cnrty-Ind R&D INTENSITY ∗

relative to OECD avg.

1,784

020

057

−096

269

Based on eight years of data for 48 contignuous states and Washington, D.C. (less ID, ND, and SD).

We leave several of these measures of technical activity unscaled, not converting them into intensities.13 We expect that there is a strong fixed cost component to technical activity—a minimum threshold exists for technical innovation. Therefore, what matters is the overall magnitude of the technical activity, not the relative intensity. For example, the District of Columbia has a relatively high technical intensity because of its small population, but its overall magnitude is small. We expect firms to be attracted to states that have surmounted these minimum thresholds of accumulating technical activity. Descriptive statistics of these state characteristics are shown in Table 1. In addition to state variables, we also need firm traits that are likely to introduce heterogeneity into valuation of state characteristics. We measure firm 13

Of the four measures R&D intensity is a scaled measure (total R&D spending/state gross product). We include R&D intensity instead of total R&D spending because R&D intensity is a standard measure in other studies. Our four measures provide a mix of both the standard technical intensity measure and several new unscaled measures of magnitude.

Management Science/Vol. 48, No. 12, December 2002

technological capabilities using R&D intensity (R&D spending scaled by sales) from the entrants’ home country-industry. These data, available at a mixture of two-digit and three-digit ISICs for 29 member nations of the OECD on a country-industry-year basis, come from the Organization for Economic Cooperation and Development (OECD) report of Main Industrial Indicators. We scale R&D intensity by the OECD average for the corresponding industry-year because we want relative technical capabilities; there are vast differences in the amount of spending between industries. Pharmaceutical and chemical industries spend more than other industries, and we need to scale these absolute values so that they do not drive the subsequent estimation. From this scaled value we subtract 1.0 so that positive (negative) values indicate that the countryindustry is above (below) average. In addition, we also collect firm-specific R&D intensity for those firms listed on the Worldscope: Global database. Because this covers only about 35% of firms in our sample, we rely primarily on the country-industry measures in our subsequent analysis; though results using the 1543

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

35% subsample yield similar results, which we discuss in the robustness section. We also include several dummy variables for those industries that have the highest R&D intensity (R&D spending/sales) within our data, those over 5% for OECD nations. We include these variables to reflect those industries that are the most knowledge intensive and, thus, where knowledge seeking might be more prevalent. We group these industries into pharmaceuticals (SIC 283), semiconductors (SIC 367), chemicals (SIC 281, 286, 287, 289), and electronics/electrical equipment (SIC 361 to 368, excluding 367).14 We group several industries into the same categories of “chemicals” and “electronics/electrical equipment” to keep the number of independent variables tractable.

Results

Our empirical approach consists of two steps. First, we identify what state traits are differentially valued by firms, e.g., what variables present coefficients with significant standard deviation in a model where all coefficients are allowed to vary randomly. Second, we explain the causes of significant standard deviations by introducing firm traits. We use the traditional conditional logit model (CLM) as an initial benchmark for the random parameter logit (RPL). To start, we introduce the set of state characteristics and allow their coefficients to vary randomly. This baseline RPL specification is shown in columns (2a) and (2b) Table 2. 14

These categories are similar to those that Florida (1997) and Serapio and Dalton (1999) find compose most of the R&D labs in the United States, less automotive, which is not as R&D intensive as the other industries. Instead of these dummies, we considered adding a continuous variable for industries’ mean R&D intensity, but this measure (OECD average R&D intensity) already appears as the denominator in the “firm” measure; the resulting correlation would likely cause artificial statistical significance. We also consider introducing additional dummy variables for other industries with R&D intensities slightly below 5%, but find that the set of dummies for industries with greater than 5% intensities explains significance in how investment firms heterogeneously value states’ R&D intensity. Introducing additional dummies for other industries is not needed to explain heterogeneity in valuation.

1544

Note that there is a column corresponding to the estimates for the means of ij (column 2a) and the standard deviations of ij (column 2b). While numerous means are significant, only two of the independent variables have significant standard deviations: average weekly wage and state R&D intensity. This suggests that of the numerous state characteristics only these two vary randomly and may be affected by firm heterogeneity. Those variables that have significant means in the RPL model also have significant coefficient estimates in the corresponding CLM model, shown as column 1 of Table 2. The improvement in fit of the RPL model (column 2) versus the corresponding CLM (column 1), measured by a likelihood ratio test, is far from significant (column 2 versus 1 at the 99.0% level). The RPL model in column 2 essentially introduces 14 unnecessary degrees of freedom for the standard deviations of coefficients because their inclusion does not substantially improve model fit. This baseline RPL model in column 2 allows us to identify those state characteristics that do and do not vary randomly and therefore where we should focus further attention when introducing firm heterogeneity. We continue to allow these two state characteristics to vary while fixing the remaining 14 independent variables. This specification is shown in columns (3a) and (3b) of Table 2. Not surprisingly, the results in column 3 are almost identical to those shown in column 2; but the likelihood ratio tests shows that this RPL specification improves the model fit versus the corresponding CLM specification (at the 10.5% level for column 3 versus 1). Allowing these two variables to vary randomly is a marginally better model specification than fixing them all to the same coefficient. The results across all columns of Table 2 indicate that states with greater LAND area are more likely to attract inward FDI, which is consistent with a dartboard approach—all other things being equal, more entrants will fall in larger areas. Further, entering firms prefer states with lower population DENSITY, where people have high INCOMEs, pay low TAXES per capita, have fewer HIGHWAY miles per capita, have a higher percentage in manufacturing, receive a higher weekly WAGE, have right-to-work laws, Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Table 2

Determinants of Inward FDI Location Choice in the United States 1987–1993 CLM betas (1)

State’s LAND area Population DENSITY Per capita INCOME Tax per capita AIRPORTS per capita HIGHWAY miles per capita % of pop employed in MANUF % UNEMPLOYMENT Average weekly WAGE Has right-to-work laws % UNIONS TAX as % of income Count of doctorates in Sci. & Eng. Doctorates EARNED in Sci. & Eng. State R&D INTENSITY PATENTS awarded residents

00095∗∗∗ 0001 −00711∗∗ 0032 01344∗∗∗ 0029 −28210∗∗∗ 0818 −35337∗ 1982 −257565∗∗∗ 6298 170750∗∗∗ 1692 00373 0027 47329∗∗∗ 1065 04690∗∗∗ 0106 00121∗∗∗ 0004 04665∗∗∗ 0153 00785∗∗∗ 0012 00416 0177 −00153∗∗∗ 0003 −00003∗∗∗ 0000

Number of observations 1784 Log-likelihood −57855 R-squared 0153 Adjusted R-squared 0153 Likelihood ratio test: Difference in log-likelihood Chi-square Prob. (current vs. (1)) Degrees of freedom

RPL all variables random means (2a) 00093∗∗∗ 0001 −00818∗∗ 0033 01545∗∗∗ 0030 −30795∗∗∗ 0832 −27113 2030 −276637∗∗∗ 6691 171277∗∗∗ 1828 00274 0028 47610∗∗∗ 1063 02587 0204 00110∗∗ 0004 05069∗∗∗ 0156 00866∗∗∗ 0013 −00042 0182 −00208∗∗∗ 0004 −00003∗∗∗ 0000

std. dev. (2b) 00005 0003 00047 0041 00115 0036 02628 0330 26726 3125 29000 8174 24645 4825 00038 0046 48184∗∗ 2017 07548 0526 00018 0007 00355 0090 00013 0004 00332 0078 00246∗∗∗ 0008 00000 0000

RPL select variables random means (3a) 00094∗∗∗ 0001 −00829 0032 01566∗∗∗ 0030 −30365∗∗∗ 0822 −25905 1997 −272377∗∗∗ 6457 172872∗∗∗ 1843 002676 0028 47885∗∗∗ 1057 03964∗∗∗ 0120 00110∗∗ 0004 04997∗∗∗ 0154 00869∗∗∗ 0013 −00109 0181 −00217∗∗∗ 0004 −00003∗∗∗ 0000

std. dev. (3b)

44253∗∗ 2237

00271∗∗∗ 0007

1784 −57797 0154 0153

1784 −57810 0154 0153

58 0990 16

45∗ 0105 2

∗ ∗∗ ∗∗∗   significant at 10%, 5%, and 1% level for 2-tailed tests. Standard errors below estimates. Note. Conditional logit (CLM) and random parameter logit (RPL) models of location choice among 45 States by inward FDI. RPL assumes coefficients vary for each chooser, and provides both the mean and std. deviation of these varying coefficients. For RPL all coefficients are assumed to be normally distributed. RPL uses 100 simulated log-likelihood functions.

Management Science/Vol. 48, No. 12, December 2002

1545

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

and have high TAX as percent of income. The positive coefficient for TAX as percent of income results from a 0.91 correlation with TAXES per capita; if only one or the other is included, either coefficient is negative as expected and significant at a 5% level; high taxes make a location less attractive. Preference for fewer HIGHWAY miles per capita reflects highdensity states that are proximate to other high-density states. States with the highest HIGHWAY miles per capita are sparsely populated and isolated, for example, Wyoming, Montana, and Nevada; while the states with the lowest AIRPORTS and HIGHWAY miles per capita are Rhode Island, Massachusetts, and Connecticut. Overall, these findings are generally consistent with prior studies and with our expectations of FDI being drawn to locations where the likelihood of high revenues, lower costs, and access to proximate markets are better.15 Turning to indicators of technical activity, interestingly, three of the measures are significant. Foreign entrants are drawn to states with high counts of doctorates in science and engineering, lower R&D INTENSITY, and that award fewer PATENTS. Depending on how strongly one believes that knowledge seeking is an important motive for pursuing FDI, we would expect either positive estimates or estimates not significantly different from zero for the means (in RPL) and betas (in CLM). The positive coefficient on count of doctorates suggests that either knowledge seeking is a tremendously frequent motive or that more likely most firms, regardless of their motives, value the presence of skilled individuals. When a firm expands motivated by internalization, it still needs to hire engineers and scientists locally, and the location of Ph.D. and non-Ph.D. scientists and engineers are likely to be strongly correlated. In contrast, the negative coefficients on PATENTS and R&D INTENSITY are initially surprising. However, recognizing that the count of patents increases not only 15

Any differences between our and Coughlin et al.’s (1991) results are attributed to our scaling by population versus CTA scaling by land area. In other specifications not reported here, we much more closely replicate CTA’s results when we also scale by land area. We prefer to scale by population because scaling by land creates several correlated proxies from population density.

1546

with greater technical activity but also with population size, this negative finding more likely suggests a preference for states with smaller populations. The parameter estimates for the mean and standard deviation of R&D INTENSITY are particularly interesting. Although the negative mean (−0
0217) implies that most firms negatively value state R&D intensity, the large variation (standard deviation equal to 0.02713) causes the coefficient to be positive for about 20% of the observations.16 In other words, while most firms are repelled by state R&D intensity, a minority finds it attractive. This heterogeneous valuation leads us to explore what type of firm values R&D intensive states. We introduce R&D intensity in the firm’s home country-industry year as a measure of firm heterogeneity that accounts for firm-specific offsets from the mean of ij . These offsets are conceptually similar to interaction variables. There are two interaction variables, one for each variable that shows coefficients with significant standard deviations in the base model. The results from introducing these interaction variables are shown in columns (1a) and (1b) in Table 3. Both interaction variables are significant, and their inclusion substantially improves model fit. Interpreting the coefficients, “weekly WAGE * Cntry-Ind R&D INTENSITY” expresses how a firm’s home countryindustry R&D intensity moves the firm away from the overall estimated mean of “weekly WAGE.” The positive parameter estimate for “weekly WAGE * Cntry-Ind R&D INTENSITY” suggests that firms from country-industries that have higher R&D intensity value a higher weekly wage even more than the average firm. This is consistent with technically advanced firms needing technical advanced employees who earn a higher wage. The other significant parameter estimate provides the offset from the mean valuation of state technical activity, R&D INTENSITY. The negative estimate for “R&D INTENSITY * Cntry-Ind R&D INTENSITY” 16

Recall that variation is assumed to follow a normal distribution. These parameter estimates describe a normal distribution centered at −0
0217 with a right tail extending into the positive numbers such that 20% are greater than zero.

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Table 3

Effect of Investments Attributes on Location Choice in the United States 1987–1993 RPL means (1a)

State’s LAND area Population DENSITY Per capita INCOME TAX per capita AIRPORTS per capita HIGHWAY miles per capita % of pop employed in MANUF % UNEMPLOYMENT Average weekly WAGE Has right-to-work laws % UNIONS TAX as % of income Count of doctorates in Sci. & Eng. Doctorates EARNED in Sci. & Eng. State R&D INTENSITY PATENTS awarded residents Weekly WAGE ∗ Cntry-Ind R&D INT R&D INT ∗ Cntry-Ind R&D INT R&D INT ∗ In Semiconductor Ind.

00091∗∗∗ −00723 01472∗∗∗ −26953∗∗∗ −32574 −247596∗∗∗ 187978∗∗∗ 00352 44282∗∗∗ 03599∗∗∗ 00105∗∗ 04254∗∗∗ 00801∗∗∗ 00655 −00220∗∗∗ −00003∗∗∗ 22470∗∗ 1017 −00117∗∗∗ 0004

std. dev. (1b)

49773∗∗

00244∗∗∗

R&D INT ∗ In Electronics Ind. R&D INT ∗ In Chemical Ind. R&D INT ∗ In Pharmaceutical Ind. Number of observations Log-likelihood R-squared Adjusted R-squared Likelihood ratio test: Difference in log-likelihood Chi-square Prob (vs prior Table 2) Degress of freedom

means (2a) 00097∗∗∗ −00755 01417∗∗∗ −25830 −39131∗ −244441∗∗∗ 192202∗∗∗ 00403 52262∗∗∗ 04489∗∗∗ 00127∗∗ 04126∗∗ 00738∗∗∗ 00278 −00261∗∗∗ −00003∗∗∗ 16895 1072 −00050 0004 00247∗∗∗ 0010 00151∗∗ 0007 −00212∗∗ 0009 00530∗∗ 0011

std. dev. (2b)

24765

00132

1784 −52043 0150 0150

1784 −51753 0155 0155

5767∗∗∗ 0000 2

6057∗∗∗ 0000 10

∗ ∗∗ ∗∗∗   significant at 10%, 5%, and 1% level for 2-tailed tests. Standard errors below estimates, some standard errors not shown in the interest of space. Not shown, four coefficients for WAGE*four industry dummies. Note. Random parameter logit (RPL) models of location choice among 45 states by inward FDI. RPL assumes coefficients vary for each chooser, and provides both the mean and std. deviation of these varying coefficients. For RPL all coefficients are assumed to be normally distributed. RPL uses 100 simulated log-likelihood functions.

indicates that firms from above average R&D intensity country-industries have more negative valuations of state R&D intensity; a negative value is added to the negative overall mean (−0
0220), yielding a larger negative. In contrast, firms from below average R&D intensity country-industries have less negative or positive valuations. Below-average firms have negManagement Science/Vol. 48, No. 12, December 2002

ative values for country-industry R&D intensity; this negative multiplied by the negative coefficient estimate for “R&D INTENSITY * Cntry-Ind R&D INTENSITY” is a positive. This positive is added to −0
0220; when country-industry R&D intensity is low enough, the overall valuation is positive. This suggests that firms from country-industries that have below aver1547

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

age R&D intensity are attracted to states with greater R&D intensities, which is consistent with our expectations for knowledge-seeking behavior by technical laggards. To further explore the heterogeneity in how investing firms value state R&D intensity, we also interact it with several industry group dummy variables for pharmaceutical, semiconductor, chemicals, and electrical/electronics, expecting that knowledge-seeking behavior is more prevalent in these R&D intensive industries. In columns (2a) and (2b) of Table 3, we see that all four of the interaction terms between state R&D intensity and the industry dummies are significant. “R&D INT * In Semiconductor Ind.” attracts a positive coefficient estimate, as do electronics and pharmaceuticals; investments into these industries more positively value state R&D intensity. Chemical industries attract a negative coefficient. Three interesting findings emerge from these offset terms. First, the coefficient attracted by the “R&D INT * In Pharmaceutical Ind.” has the largest positive valuation, suggesting that knowledge seeking is more important in pharmaceuticals than in any other industry. This positive valuation is over twice that of firms entering the semiconductor industry (0.0530 versus 0.0247) and four time that of firms entering other electronics industries (0.0530 versus 0.0151). Second, knowledge seeking is limited to only knowledge-intensive industries. When these dummy offset variables are included, the standard deviation estimate for state R&D intensity becomes nonsignificant (“State R&D INTENSITY” in column (1b) versus (2b). This suggests that including these measures of firm heterogeneity explains much of the underlying variation in how firms value state R&D intensity. Third, knowledge seeking occurs not only among laggards, but leaders as well. To support this claim, we obtain the R&D intensities for the four industry groups studied. Investments in semiconductors, electronics, chemicals, and pharmaceuticals have R&D intensities that are 42.4% below, 38.4% below, 5.7% above, and 14.1% above average.17 Firms investing 17

A raw country-industry R&D intensity is scaled by the OECD industry average to normalize differences across industries. This normalized value is compared to the average R&D intensity of

1548

in semiconductors and electronics are below average and more positively value state R&D intensity. Conversely, firms investing in chemicals are above average and more negatively value state R&D intensity. These three groups conform to “technical laggards seek knowledge to catch up, while leaders do not.” Results for the pharmaceutical industry are different. Firms investing in pharmaceuticals are above average and more positively value state R&D intensity. This suggests that technical leaders invest in the United States and seek out locations of greater technical activity. Why this different strategic behavior in pharmaceuticals? One potential explanation is the specific market requirements for selling pharmaceuticals in the United States; FDA approval is needed and accessing specialized suppliers to handle FDA trials might affect foreign firms’ location choice.18 Another explanation is suggested by Cantwell and Janne (1999): that firms from leading geographic centers may go abroad not to catch up, but to source greater technical variety. With discrete product innovation, high technical uncertainty, and long development times, all firms in the pharmaceutical industry may be forced to locate in R&D intensive locations to monitor competitors’ advances and to identify and test new emerging compounds. Consistent with pharmaceutical’s uniqueness, Lim (2001) finds key differences for innovation in the pharmaceutical and semiconductor industries; while basic research is geographically concentrated in both, semiconductors innovations occur widely, suggesting that innovation is not tightly tied to basic research. In contrast, pharmaceutical innovations are linked to basic research, which suggests the importance of collocating where basic research occurs. investing firms in the United States to determine extent of over or under average. For example, for pharmaceuticals investing firms have 14.7% R&D intensities. This is scaled by the OECD average of 11.0% for pharmaceuticals. From 1.341 we subtract 1.201 (firms investing in the United States come from country-industries that have R&D intensities 20.1% greater than the OECD average for that industry), which is 0.141 or 14.1%. 18

The authors thank Lorraine Eden and an anonymous reviewer for this argument.

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Overall, our results suggest that a sizeable minority of investment into the United States is seeking technical knowledge. This behavior occurs predominantly in knowledge-intensive industries—those with high R&D intensities. Within these high R&D industries, most firms that come from country-industries with below average R&D intensity are attracted to states with greater R&D intensities. The exception is investments in the pharmaceutical area, where technical leaders seek out locations of greater technical activity. Finally, firms investing in pharmaceuticals value R&D intensive states several times more than do firms investing in other industries.

Robustness Checks

Before settling on these results, we conduct several alternate tests. First, while using country-industry R&D measures, firm-specific R&D measures are available for about 35% of the investment transactions in our sample. We repeat several key model specifications using this subsample. The results are in Table 4. Table 4 has two model specifications. The first has select variables random with R&D intensity interaction terms, which parallels Table 3, columns (1a) and (1b). The second model parallels Table 3, columns (2a) and (2b) by adding industry dummy interaction terms. The results for this subsample and the full sample share several common findings. For both, the state R&D intensity mean attracts a negative coefficient. For both, the coefficient estimate for the standard deviation of state R&D intensity is significant and large enough to suggest that while most firms are repelled, that a minority are attracted by states with greater technical intensity. For the interaction term of state R&D INT and firm R&D INT, the coefficient is negative but not significant for the subsample. Both sets of results suggest that knowledge seeking is limited to knowledge-intensive industries; industry interaction dummy variables are significant and their inclusion robs the state R&D intensity standard deviation estimate of statistical significance. Also, the magnitude of coefficients attracted by the industry offset terms are similar, suggesting that firms investing in pharmaceuticals have the greatest positive valuation, followed by semiconductors and electronics. Thus, most of our main findings are supported in this subsample. Management Science/Vol. 48, No. 12, December 2002

Second, many of the 1,784 transactions may come from the same firms because firms might invest multiple times during our 1987–1993 period. If so, the location choice of such observations are likely not independent; where a foreign firm has already invested will affect where it subsequently invests. To address this potential lack of independence we examine the declared foreign parent for each transaction to identify those parents that invested more than once. By matching names we find that each foreign parent made, on average, 1.57 investments during our study period. We repeat the specifications shown above using only those parents that made only one investment during our sample period. With this reduced sample of 754 observations the results are almost identical to the specifications shown in Table 3 in sign, magnitude, and significance. Third, we expect that mode of entry does not constrain location choice: If a desired location does not contain a suitable acquisition target, then the firm can build a new greenfield facility. While including acquisitions in prior tests, we explore whether excluding acquisitions affects our results. We exclude acquisitions, which is about half of the observations, and repeat the test specifications found in Tables 2 and 3. Our results are less significant, potentially due to having fewer observations, but qualitatively similar in sign and magnitude to those reported previously. Finally, a state level of analysis may be inappropriate for testing some of our expectations. While costminimizing firms may choose states with low taxes or right-to-work laws, knowledge seeking likely follows other boundaries. States may be too narrow or too broad; firms may locate in New Jersey to access New York City, while Texas has several potential economic centers. Therefore, we also conduct an analysis using the 170 Economic Areas (EAs) defined by the Bureau of Economic Analysis (BEA) as shown in Figure 2. EAs are defined to maximize the percent of people that live and work within the same area. While EAs provide a more meaningful definition of locations based on economic boundaries, they have an important shortcoming: lack of data at the EA level. Since the BEA reports only five data series, we construct our other variables from the component state 1549

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Table 4

Determinants of Inward FDI Location Choice in the United States 1987–1993 (Firm R&D) Firm R&D subsample means (1a)

std. dev. (1b)

means (2a)

State’s LAND area Population DENSITY Per capita INCOME

00111∗∗∗ −00609 01792∗∗∗

00116∗∗∗ −00687

TAX per capita

−50188∗∗∗

−48930∗∗∗

AIRPORTS per capita HIGHWAY miles per capita % of pop employed in MANUF % UNEMPLOYMENT Average weekly WAGE Has right-to-work laws % UNIONS TAX as % of income Count of doctorates in Sci. & Eng.

01826∗∗∗

−50200 −207992 224078∗∗∗ 00949∗ 65768∗∗∗

std. dev. (2b)

−56215 −194926 224377∗∗∗ 00944∗ 02089

02095 00044 09336∗∗∗

89002∗∗∗

03561

02514 00058 09127∗∗∗

01036∗∗∗

00999∗∗∗

Doctorates EARNED in Sci. & Eng. State R&D INTENSITY

01010 −00218∗∗

−01499

PATENTS awarded residents

−00004∗∗

−00003∗∗

WAGE ∗ Firm R&D INT

−02633 0528

−05507 0466

R&D INT∗ Firm R&D INT

−00021 0003

−00005 0003

00256∗

R&D INT∗ In Semiconductor Ind.

−00310∗∗∗

00176

00339∗∗ 0041 −00232 0014

R&D INT∗ In Chemical Ind. R&D INT∗ In Electronics Ind.

00228∗∗

R&D INT∗ In Phamrmaceutical Ind.

00634∗∗

0011 0016 Number of observations Log-likelihood R-squared Adjusted R-squared

626 −19668 0179 0179

626 −19503 0186 0185

∗ ∗∗ ∗∗∗   significant at 10%, 5%, and 1% level for 2-tailed tests. Standard errors below estimates, some standard errors not shown in the interest of space. Not shown, four coefficients for WAGE∗ four industry dummies. Note. Random parameter logit (RPL) models of location choice among 45 states by inward FDI. RPL assumes coefficients vary for each chooser, and provides both the mean and std. error of these varying coefficients. For RPL all coefficients are assumed to be normally distributed. RPL uses 100 simulated log-likelihood functions.

1550

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

data.19 We split the full 170-choice set into four subsets due to limitations of the statistical package and then test the main specifications on these four choice sets.20 The results are shown in Table 5. In the interests of space we show results only from choice set 1 (results for choice sets 2 and 4 are similar to set 1) and choice set 3. The results for these two sets are qualitatively similar to the state-level analysis though a difference in model fit stands out. The EA-level analysis explains substantially more variation. Depending upon the choice set, the models explain from a low of 19.6% to a high of 40.2% (for choice set 2, not shown) compared to 15.5% for the various analyses in Tables 2 and 3. This difference suggests that the EA-based analysis is superior. Interestingly, only choice set 3 has a significant negative coefficient for the interaction term of EA and firm country-industry R&D intensity, which suggests that knowledge seeking only occurs in specific regions within the United States. Of note, by chance, set 3 includes Boston (EA3) and San Francisco (EA163), both which are clearly technology centers. It also includes several other potentially important centers, such as Raleigh-Durham-Chapel Hill (EA19), DallasFort Worth (EA127), and Houston-Galveston (EA131).

Conclusions

We examine state and firm traits that determine where inward foreign direct investments in manufac19

The BEA reports personal income, population, per capita income, total earnings, and total employment (though employment in manufacturing is broken out). To construct other variables, we use population weighted averages of the component states’ values. Conceptually, if a region is 70% New York, 30% New Jersey, 5% Connecticut, and 5% Pennsylvania, then the region’s value is an average of these contributing states’ values, weighted by their percents (we use percents based upon populations—if 70% of an area’s population is NY, 30% NJ, etc.).

20

In splitting the choice sets, we are conscious that EA identification numbers are assigned from east to west, starting in Maine, and winding back and forth in a serpentine fashion from north to south. Low numbers are along the East Coast and high ones are on the West Coast. Wanting choice sets that randomly represent the entire nation, we use sets that are composed of every 4th area. Specifically, choice set 1 is composed of areas 1, 5, 9, 13, 17


169; choice set 2 is areas 2, 6, 10, 14, 18


170; choice set 3 is areas 3, 7, 11, 15, 19


167; choice set 4 is areas 4, 8, 12, 16, 20


168.

Management Science/Vol. 48, No. 12, December 2002

turing from OECD nations locate within the United States for 1987–1993. Because firms are motivated to conduct FDI for numerous reasons, we use the random parameter logistic regression (RPL) to make use of this heterogeneity. The RPL allows the effect of state characteristics to vary randomly for each investing firm, versus the more traditional conditional logit model (CLM), which constrains the effect of state characteristics to be the same for all firms. The RPL allows each firm to have a different valuation. We then capture this variation in valuation resulting from heterogeneous motives by including additional characteristics for each investment, such as what industry the investing firm is in. Besides confirming prior results that showed that states with greater market size, lower factor costs, and better access to surrounding states attract more FDI, our analysis provide three new findings that contribute to our understanding of knowledge seeking via FDI. First, knowledge seeking is limited to firms in research-intensive industries. Although state R&D intensity does not attract FDI on average, a significant minority of firms—mostly in research-intensive industries—is attracted to states with high R&D intensity. Also interesting is that we find this result among investments whose primary function is manufacturing. This suggests that participants in knowledgeintensive industries seek knowledge not only through laboratories but also via manufacturing facilities and is consistent with Patel and Vega (1999), who find that innovations of foreign firms abroad are also more applied; many patents of foreign firms abroad are for processes and machinery. Second, we are able to quantify differences in how important the technical knowledge available in a location is for investing firms. While prior studies demonstrate that knowledge seeking is important for various industries, our findings suggest an ordering of importance. We find that investing firms in the pharmaceutical industry value state R&D intensity the most. This positive valuation is over twice that of firms entering the semiconductor industry and four times that of firms entering other electronics industries. Third, knowledge seeking occurs not only among technical laggards, but also among technically leading 1551

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Table 5

Determinants of Inward FDI Location Choice in the United States 1987–1993 (Economic Areas) Choice set 1 EAs: 1 5 9    169 means (1a)

Region’s LAND area Population DENSITY Per capita INCOME TAX per capita AIRPORTS per capita HIGHWAY miles per capita % of pop in MANUF % UNEMPLOYMENT Average weekly WAGE Has right-to-work laws % UNIONS TAX as % of income Econ area R&D INTENSITY

0060∗∗∗ 001 18888∗∗∗ 357 0170∗∗ 007 −0005∗∗ 000 −8657 853 −0026 002 0146∗∗∗ 004 −0063 010 −0008∗∗ 000 −1330 123 0057∗∗∗ 001 0932∗∗ 047 −0028∗∗ 001

std.dev. (1b) 0005 001 0359 436 0083 015 0000 000 9047∗∗ 381 0005 001 0013 006 0083 021 0000 001 1807 164 0002 003 0057 027 0066∗∗ 003

Choice set 3 EAs: 3 7 11    167 means (3a) 0014∗∗∗ 001 5797∗∗∗ 208 0326∗∗∗ 005 −0006∗∗∗ 000 −7808∗∗ 381 −0065∗∗∗ 001 0155∗∗∗ 003 0023 005 0007∗∗∗ 000 −1456∗∗∗ 027 −0044∗∗∗ 001 1129∗∗∗ 024 −0023∗∗ 001

std. dev. (3b) 0001 000 1096 352 0017 006 0000 000 8453∗∗∗ 241 0004 002 0016 003 0050 011 0007∗ 000 0069 053 0017 002 0027 017 0032∗∗∗ 001

R&D INT ∗ Cntry-Ind R&D INT Number of observations Log-likelihood R-squared Adjusted R-squared Chi-squared

260 −6487 0272 0270 48549∗∗∗

603 −16489 0196 0196 80621∗∗∗

Choice set 1 EAs: 1 5 9    169 means (1c) 0058∗∗∗ 001 20832∗∗∗ 354 0174∗∗ 009 −0004∗ 000 0942 643 −0025 003 0154∗∗∗ 004 −0047 012 −0008∗∗ 000 −0202 056 0065∗∗∗ 001 0854∗ 047 −0011 002 −0014 001

std. dev. (1d)

0071∗∗ 003

260 −5924 0271 0270 44002∗∗∗

Choice set 3 EAs: 3 7 11    167 means (3c) 0017 001 −4992∗∗ 197 0297∗∗∗ 004 −0006∗∗∗ 000 −6396∗ 362 0066∗∗∗ 001 0168∗∗∗ 003 0022 005 0007∗∗∗ 000 −1201∗∗∗ 024 −0039∗∗∗ 001 1112∗∗∗ 022 0004 001 −0022∗∗∗ 000

std. dev. (3d)

0034∗∗∗ 001

603 −15337 0189 0188 71505∗∗∗

∗ ∗∗ ∗∗∗   significant at 10%, 5%, and 1% level for 2-tailed tests. Standard errors below estimates. Note. Random parameter logit (RPL) models of location choice among 170 economic areas (EA) buy inward FDI. Due to statistical packages choice set limit, 170 economic areas are split into four smaller choice sets are composed of each 4th EA. Only the 1st and 3rd sets are shown. RPl assumes coefficients vary for each other chooser, and provides both the mean and std. error of these varying coefficients. For RPL all coefficients are assumed to be normally distributed. RPL uses 100 simulated log-likelihood functions.

1552

Management Science/Vol. 48, No. 12, December 2002

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

firms. We find that firms investing in the pharmaceutical industry positively value state R&D intensity even though they come from technically leading nations, whereas firms in other knowledge-intensive industries that also positively value state R&D intensity originate mostly from technically lagging nations. Prior research that focused on differences in R&D— on R&D laggards—might overlook situations such as this. Certain industries are based upon technical competition, which may force all participants to seek spillovers from competitors. This differential outcome of laggards and some leaders seeking knowledge is consistent with Cantwell and Janne (1999), who find that firms originating from leading versus lagging technical locations establish subsidiaries for different strategic purposes. Overall, while refining our understanding of knowledge-seeking behavior, our investigation highlights the importance of firm heterogeneity in international strategy. By examining whether and how strongly firms are attracted to the locations’ technical activity, we find two levels of firm heterogeneity: (i) firms invest abroad for different reasons—most prominently to internalize existing capabilities and to seek new knowledge, and (ii) within knowledge seeking, firms have different motivations—laggards catching up and leaders sourcing technical diversity. More research is needed to fully understand knowledge seeking. Besides heterogeneity on the demand side (heterogeneity in firms’ motives), there is also variation to be explored on the supply side (heterogeneity in creators of new knowledge). Technical activity relevant for each industry resides not only in different locations, but emanates from different sources: academic, industry, and government. With more detailed data and powerful statistical tools available, further exploration of underlying heterogeneity in knowledge supply should yield additional insights into the knowledge-seeking process. Acknowledgments The authors acknowledge the helpful comments of two anonymous reviewers and the associate editor, Lorraine Eden, Mauro Guillen, Witold Henisz, Glenn Hoetker, Arturs Kalnins, Andy King, Joanne Oxley, Tom Pugel, Myles Shaver, and seminar participants at the New York University Multinational Conference on prior drafts.

Management Science/Vol. 48, No. 12, December 2002

Appendix

Variable Definitions and Sources

State characteristics Coverage

Source

State LAND area

1987–1993 General Services Administration Population DENSITY 1987–1993 Census Bureau (population) Per capita INCOME 1987–1993 Bureau of Economic Analysis TAX per capita 1987–1993 Bureau of Economic Analysis AIRPORTS per capita 1987–1993 Department of Transportation HIGHWAY miles per capita 1987–1993 Department of Transportation % of pop employed in MANUF 1987–1993 Bureau of Labor Statistics % UNEMPLOYMENT 1987–1993 Bureau of Labor Statistics Average weekly WAGE 1987–1993 Bureau of Labor Statistics Has right-to-work laws 1987–1993 National Right to Work Legal Defense Foundation % UNIONS 1987–1989 Manufacturing Climates Study TAX as % of income 1987–1993 Bureau of Economic Analysis Count of PHDs in Sci & Eng 1995 National Science Foundation PHDs EARNED in Sci & Eng 1995 National Science Foundation R&D INTENSITY (%) 1995 National Science Foundation PATENTS awarded residents 1995 National Science Foundation FDI transaction characteristics Coverage Cntry–Ind R&D INTENSITY

Source

1987–1993 OECD—Main Industrial Indicators

References

Almeida, Paul. 1996. Knowledge sourcing by foreign multinationals: patent citation analysis in the US semiconductor industry. Strategic Management J. 17(Winter) 155–165. , Bruce Kogut. 1999. Localization of knowledge and the mobility of engineers in regional networks. Management Sci. 45(7) 905–917. Berry, Steve, James Levinsohn, Ariel Pakes. 1995. Automobile prices in market equilibrium. Econometrica 60(4) 889–917. Blonigen, Bruce A. 1997. Firm-specific assets and the link between exchange rates and foreign direct investment. Amer. Econom. Rev. 87(3) 447–465. Buckley, Peter J., Mark C. Casson. 1976. The Economic Theory of the Multinational Enterprise. Macmillan, London, U.K. Cantwell, John. 1989. Technological Innovation and Multinational Corporations. Basil Blackwell, Oxford, U.K. , Odile Janne. 1999. Technological globalization and innovation centers: The role of corporate technological leadership and locational hierarchy. Res. Policy 28(2, 3) 119–144. Caves, Richard E. 1971. International corporations: The industrial economics of foreign investment. Economica 38 176–193. . 1996. Multinational Enterprise and Economic Analysis, 2nd ed. Cambridge University Press, Cambridge, U.K. Coase, R. H. 1937. The nature of the firm. Economica 4(16) 386–405.

1553

CHUNG AND ALCÁCER Knowledge Seeking and Location Choice

Coughlin, C., J. Terza, V. Arromdee. 1991. State characteristics and the location of foreign direct investment within the United States. Rev. Econom. Statist. 73 675–683. Florida, Richard. 1997. The globalization of R&D: Results of a survey of foreign-affiliated R&D laboratories in the USA. Res. Policy 26 85–103. Greene, William. 2000. Econometric Analysis, 4th ed. Prentice-Hall, Upper Saddle River, NJ. Hennart, Jean-Francois, Young-Ryeol Park. 1993. Greenfield vs. acquisition: The strategy of Japanese investors in the United States. Management Sci. 39(9) 1054–1070. Jaffe, Adam B., Manuel Trajtenberg, Rebecca Henderson. 1993. Geographic localization of knowledge spillovers as evidenced by patent citations. Quart. J. Econom. 108(3) 577–598. Kogut, Bruce, Sea Jin Chang. 1991. Technological capabilities and Japanese foreign direct investment in the United States. Rev. Econom. Statist. 73(3) 401–413. , Udo Zander. 1992. Knowledge of the firm, combinative capabilities, and the replication of technology. Organ. Sci. 3 383–397. Kuemmerle, Walter. 1999. The drivers of foreign direct investment into research and development: An empirical investigation. J. Internat. Bus. Stud. 30(1) 1–24. Lim, Kwanghui. 2001. The relationship between research and innovation in the semiconductor and pharmaceutical industries (1981–1997). Res. Policy Forthcoming.

Morck, Randall, Bernard Yeung. 1991. Why investors value multinationality. J. Bus. 64(2) 165–187. Patel, Pari, Modesto Vega. 1999. Patterns of internationalization of corporate technology: Location vs. home country advantages. Res. Policy 28(2, 3) 145–155. Pugel, Thomas, Erik Kragas, Yui Kimura. 1996. Further evidence on Japanese direct investment in U.S. manufacturing. Rev. Econom. Statist. 78(2) 208–213. Revelt, David, Keith Train. 1998. Mixed logit with repeated choices of appliance efficiency levels. Rev. Econom. Statist. 80(4) 647–657. Serapio, Manuel, Donald Dalton. 1999. Globalization of industrial R&D: An examination of foreign direct investment in R&D in the United States. Res. Policy 28(2, 3) 303–316. Shan, Weijian, Jaeyong Song. 1997. Foreign direct investment and the sourcing of technological advantage: Evidence from the biotechnology industry. J. Internat. Bus. Stud. 28(2) 267–284. Shaver, J. Myles. 1998. Accounting for endogeneity when assessing strategy performance: Does entry mode choice affect FDI survival? Management Sci. 44(4) 571–585. Villas-Boas, J. Miguel, Russel S. Winer. 1999. Endogeneity in brand choice models. Management Sci. 45(10) 1324–1338. Wesson, Tom. 1993. An alternative motivation for foreign direct investment. Unpublished doctoral dissertation, Harvard University, Boston, MA.

Accepted by Linda Argote; received October 2, 2000. This paper was with the authors 10 months for 3 revisions.

1554

Management Science/Vol. 48, No. 12, December 2002