Testing for causality between FDI and economic growth using heterogeneous panel data

Abstract

The causal relationship between FDI inflows and growth is of great policy interest, yet the state of concrete knowledge on the issue is rather poor. Our contribution is to investigate the causal relationship between the ratio of FDI to GDP (FDIG) and economic growth (GDPG) using a battery of cutting-edge methods and an extensive data set. We employ the heterogeneous-panel tests of the Granger non-causality hypothesis based on the works of Hurlin, C. 2004a. Testing Granger Causality in Heterogeneous Panel Data Models with Fixed Coefficients. Mimeo: University of Orléans, (Fisher, R. A. 1932. Statistical Methods for Research Workers. Edinburgh: Oliver & Boyd., Fisher, R. A. 1948. ‘Combining Independent Tests of Significance.’ American Statistician 2 (5): 30–31) and Hanck, C. 2013. ‘An intersection test for panel unit roots.’ Econometric Reviews 32 (2): 183–203. Our panel data set is compiled from 136 developed and developing countries over the 1970-2006 period. According to the Hurlin and Fisher tests, FDIG unambiguously Granger-causes GDPG for at least one country. However, the results from these tests are ambiguous regarding whether GDPG Granger-causes FDIG for at least one country. Using a test based upon Hanck, C. 2013. ‘An intersection test for panel unit roots.’ Econometric Reviews 32 (2): 183–203, both with and without one structural break in the vector autoregression, we are able to determine whether and for which countries there is Granger-causality. This test suggests that at most there are six countries (Estonia, Guyana, Poland, Switzerland, Tajikistan and Yemen) where FDIG Granger-causes GDPG and at most four countries (Dominican Republic, Gabon, Madagascar and Poland) where GDPG Granger-causes FDIG.

KEYWORDS:

• Granger-causality tests
• panel data
• FDI
• economic growth

JEL Classifications:

• C33
• F21
• F43
• N10

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 There may also exist drawbacks for the host country, e.g. a deterioration of the trade balance (the flip side of the improvement of the capital account) and crowding out of domestic investment.

2 See Chowdhury and Mavrotas (2005) and Ozturk (2007) for surveys of the FDI and growth relationship. Mody and Murshid (2002) discusses the relationship between domestic investment and FDI. See Asiedu (2003) for an excellent discussion of the relationship between policy reforms and FDI in the case of Africa. Gorg and Greenaway (2004) analyse the effects of FDI on domestic firms.

3 In our estimation we do not distinguish between developed and underdeveloped countries because Hanck’s (2013) method allows us to identify whether Granger-causality exists for each individual country.

4 The panel test statistic is not always positive, although it is based on individual Wald statistics that are all positive, because the expected value of these statistics is subtracted in constructing the normalised Z statistics. Nevertheless, the test is one-tailed because only very small values of Wald statistics will fall in the extreme left-hand tail and these will indicate non-rejection of the null. Hence, the rejection region only occurs in the right-hand tail. For extensive and full derivations of asymptotic and semi-asymptotic distributions see Hurlin (2008).

5 When the panel is balanced, and the lag lengths are the same in each cross-section’s VAR a simplified panel test statistic may be employed – see Hurlin (2008).

6 This is suggested to be true, for example, when $N=5$. This is so even when the time-series is around 50 observations, a typical size for annual macroeconomic time-series.

7 Alternatively, one can group countries into the value of $N$, $T_i$ and $H_i$ used in the test and identify the critical value appropriate for each group using (6). To obtain the critical value for the whole panel one can take the weighted average of these group critical values where the weights reflect the proportion of cross-sectional units from the whole panel appearing in each group.

8 Hurlin (2008, 11) provide the following commentary within the context of bivariate GNC tests between financial development and GDP growth. ‘What is the main advantage of this Granger non-causality test? For instance, let us assume that there is no causality from financial development to growth for all of the $N$ countries. Given the Wald statistics properties in small sample[s], the analysis based on $N$ individual tests is likely to be inconclusive. With a small $T$ sample, some of the realizations of the individual Wald statistics are likely to be superior to the asymptotic critical values of the chi-square distribution. These ‘large’ values of individual statistics lead to wrongly reject the null hypothesis of non-causality for at least some countries. The conclusions are then no[t] clear cut. On the contrary, in our panel average statistic, these “large” values of individual Wald statistics are crushed by the others which converge in probability to zero. When $N$ tends to infinity, the cross-sectional average is likely to converge to zero. The null hypothesis of [the] homogeneous non-causality hypothesis will not be rejected.’ Our comments are given in squared parentheses.

9 In being able to account for general forms of cross-sectional dependence Hanck (2013) argues that this has advantages over many second-generation panel unit root tests where non-trivial decisions are required by the user in the implementation of the tests that may affect the outcome. Such decisions are not required in the application of the intersection unit root test. Hanck (2013, 4–5) shows that the intersection test controls size for patterns of cross-sectional dependence often assumed in panel models with dynamics.

10 The procedure is appropriate for probability values based on test statistics that are multivariate totally positive of order two. This contains a large class of distributions including the absolute valued multivariate normal, absolute valued central multivariate t and central multivariate F, see Hanck (2013). Given that GNC tests can be based on t, F and chi-squared distributions this would make this an appropriate test for use with Hanck’s (2013) procedure.

11 In identifying which cross-sectional units in the panel reject the null and those which do not using a large number of tests Hommel (1988) proves that the above procedure controls for the ‘Familywise Error Rate’ (FWER). That is, in choosing the level of significance for an individual test to be $\alpha$, the above procedure ensures that the size of the test for at least one unit’s $H_{i,0}$ is $\alpha$.

12 When there are an even number of time-series observations for country $i$ both sub-samples have the same number of observations. When there are an odd number of time-series observations the first sub-sample has one more data point than the second sub-sample. $T_i^{Yr}$ denotes the date of the first observation of the second sub-sample.

13 The five countries where there were insufficient degrees of freedom to estimate VAR models that allow a break were Bangladesh, Burkina Faso, Kyrgyz Repubic, Mongolia and the Slovak Republic.