Knowledge Spillovers Through FDI and Trade: The Moderating Role of Quality-Adjusted Human Capital

Abstract

The paper extends the findings of the Coe and Helpman (Eur Econ Rev 39(5):859–887, 1995) model of R&D spillovers by considering foreign direct investment (FDI) as a channel for knowledge spillovers in addition to imports. Deeper insights on the issue are provided by examining the inter-relationship between knowledge spillovers from imports and inward FDI. Furthermore, human capital is added to the discussion as one of the appropriability factors for knowledge spillovers, with special focus on its quality-content, using journal publications and patent applications. Applying cointegration estimation method on 20 European countries from 1995 to 2010, the direct effects of FDI-related as well as import-related spillovers on domestic productivity are confirmed. Furthermore, a strong complementary relationship is found between knowledge spillovers through the channels of imports and inward FDI. When considering quality-adjusted human capital, countries with better human capital are found to benefit not only from direct productivity effects, but also from absorption and transmission of international knowledge spillovers through imports and inward FDI. Finally, technological distance with the frontier does not appear to play a role in the absorption of import and FDI related knowledge spillovers.

Appendix

Table 4 Advantages and disadvantages of different proxies for quality of human capital

Table 5 OECD Technology intensity classification

Table 6 Correlation table

1.1 A.1 Country-Wise Time Plots of Variables

Fig. 3

log(R&D Domestic)

Fig. 4

log(FDI Stock)

Fig. 5

log(Human Capital Quality)

Fig. 6

Import related spillovers

Fig. 7

log(Total Factor Productivity: Base Year = 2005)

Table 7 Estimation results: traditional human capital

1.2 A.2 Additional Estimation Results with Traditional Barro-Lee Type Human Capital Variable

1.2.1 A.2.1 Brief Overview of Cointegration

Data in macroeconomics generally possess a strong deterministic trend, especially when there is a sufficiently long time series. The variables in such cases are generally non-stationary (that is, they do not have a constant mean and variance over time). In time series, when variables are non-stationary, conventional estimation techniques, such as ordinary least squares, are expected to be driven by spurious correlation (Phillips 1986). Engle and Granger (1987) show that linear combination of two or more I(1) (non-stationary) variables could be I(0) (stationary) in which case the series are said to be cointegrated. In other words, non-stationary variables are said to be cointegrated if the residuals from their relationship are stationary. By using cointegration, one can use full information embodied in the variables and also use the attractive properties of cointegration techniques such as super consistency when n goes to infinity (Stock 1987). Estimates generated by ordinary least squares, however, do not follow an asymptotic Gaussian distribution, therefore standard testing procedures are invalid unless they are significantly modified. Fully Modified OLS (FMOLS) and Dynamic OLS (DOLS) are generally considered as an alternative to simple OLS in the presence of cointegration. Since our data contain relatively large macroeconomic time series of 16 years, we test our variables for unit root, the presence of which motivates the test for cointegration.

In time series, the Engle and Granger (1987) cointegration test is used on I(1) variables to test for cointegration. If the residuals from the regression are I(0) then the variables are said to be cointegrated. On a similar principle, Pedroni (1999, 2004) and Kao (1999) propose cointegration tests for panel data. The Pedroni test consists of several tests under different assumptions on constants and trends across cross-sections. Consider the following regression:

$$ {y}_{i,t}={\alpha}_i+{\delta}_it+{\beta}_1{x}_{1\left(i,t\right)}+{\beta}_2{x}_{2\left(i,t-1\right)}+{\beta}_M{x}_{M\left(i,t\right)}+{\varepsilon}_{i,t} $$

(10)

The variables x and y are assumed to be I(1). The individual constant and trends are represented by α and δ, respectively. The null hypothesis of the test is ‘no cointegration’. In the case of no cointegration, residuals c are integrated of order 1. If c is I(0) then the variables are said to be cointegrated. Formally, the null hypothesis of no cointegration implies ρ = 1 in Eq. (11).

$$ {\varepsilon}_{i,t}={\rho}_i{\varepsilon}_{i,t-1}+{u}_{i,t} $$

(11)

Pedroni proposes two sets of hypotheses for between and within dimension. Under the test for between dimension, the test allows for different cointegrating relationships across cross-sections, while under the test for within dimension the cointegrating relationship is assumed to be homogenous across cross sections. Eleven statistics are calculated for the Pedroni test under the assumptions described above. For the decision rule, however, there is no concrete guideline for how many tests out of eleven should show a cointegrating relationship. In this study, we reject the null of no cointegration if six out of eleven statistics of Pedroni reject the null of cointegration. Kao (1999) uses the similar approach as that of Pedroni but allows for cross section specific constants and homogenous coefficients in the first stage regressions. The null hypothesis, similar to Pedroni test, is no cointegration. For robustness of the results, we have used both Kao and Pedroni tests for cointegration.

Table 8 Cointegration tests: detailed