Is foreign direct investment attracted to ‘knowledge creators’?

Abstract

The effect of ‘knowledge creators’ on location patterns of new foreign plants entering the USA from 1986 to 1993 is analysed. The empirical results from a conditional logit model suggest a link exists between knowledge bases, measured by patent counts, and the location decisions of foreign plants. In the limit, these results imply that a 1% increase in patent counts is associated with an increase in the probability of attracting a new foreign plant by as much as 1.874%.

Notes

 The residence of the first-named inventor and technological antecedents (backward citations) are two pieces of information contained in a patent document. It is possible to trace a patent's forward citations from these. Knowledge spillovers are localized within a country (or region or SMSA) if a disproportionately high number of a patent's forward citations originate within the same country (or region or SMSA).

 Although evidence favours the existence of knowledge externalities, their effects tend to dissipate with distance (see, e.g., Jaffe, 1989; Jaffe et al., 1993; Feldman, 1994; Audretsch and Feldman, 1996). This suggests that if a plant is to take advantage of the benefits from these knowledge externalities, proximity to the source of externalities may be a necessary condition. In light of this finding, locations where knowledge spillovers are potentially large would attract more plants.

 Recognizing the role innovation can play in economic growth, all states by now have established science and technology (S&T) offices. Some have explicitly adopted S&T strategic plans emphasizing the creation of high technology plants and incorporation of new technologies into processes and products (see Feller, 1997; NSF, 1999).

 This provides the empirical analysis with a spirit similar to the treatment / control conditions in field experiments (see, e.g., List, 2001a). Also, an integer-based regression model has been experimented with (Poisson and zero-inflated Poisson – see, e.g., List, 2001b) and results are not significantly different.

 The strong assumption that the error terms ( μ  _ij ) are independent and identically distributed imposes the  ‘ Independence of Irrelevant Alternatives ’  (IIA) restriction on the predicted values. This assumption poses problems since it stretches the bounds of credulity to assume that, for example, a foreign plant's decision not to locate in Wyoming is independent of its decision to reject Idaho and Montana. This problem is mitigated by including four Census region dummies. If the error terms are only correlated within regions and not across regions, the Census dummies will capture this correlation and reduce the IIA problem. However, the equation will be mis-specified if correlation exists between states across regions.

 According to ITA's definition, FDI is  ‘ direct or indirect ownership of 10 percent or more of the voting securities of an incorporated business enterprise, or the equivalent interest in an unincorporated business enterprise. ’ 

 The information contained in the ITA publication also includes the 4-digit SIC code of the investment, the investing country, and in some cases, the amount of the completed transaction. Ideally, it would be desirable to use the total value of the investment as a function of the exogenous variables; but this information in general is available for less than 50% of the new plant investments reported in the ITA publication. Another limitation of the data is that the ITA does not track whether listed projects were completed. An occurrence though is not reported unless there are signs of completion, for example, ground breaking for new plant investment.

 However, the correlation between the spatial distribution of patent counts and actual innovation counts is quite high. Feldman and Florida (1994) report a correlation of 0.93 between patent counts and the count of innovations (for 1982) in the Small Business Administration's (SBA) Innovation Database. Also, Acs et al. (2002) obtain similar results when they use either patent count or actual innovation count as output indicator in the knowledge production function regression equation. Hence, patent count appears to be a good proxy for innovation count.

 Indeed, using unit root tests, Co and Wohar (2004) find that  ‘ technological ’  shocks have permanent effects on log relative patent per capita in some states.