Abstract
This paper contributes to the growing literature on the milex-growth nexus, by providing a case study of South Africa and considering the possibility of structural breaks by applying newly developed econometric methods. Using full sample bootstrap Granger non-causality tests, no Granger causal link is found between military expenditure and GDP for 1951–2010, but parameter instability tests show the estimated VARs to be unstable. Using a bootstrap rolling window estimation procedure, however, finds evidence of bidirectional Granger causality in various subsamples. This implies standard Granger non-causality tests, which neither account for structural breaks nor time variation may be invalid.
Notes
1 However, Batchelor, Dunne, and Saal (Citation2000b) and Birdi, Dunne, and Saal (Citation2000) considered the impact of military spending on corporate performance and industrial growth, respectively.
2 This is an extended data series from the one published and was made available by the SIPRI military spending project. We are grateful to Sam Perlo-Freeman and Mehmet Uye for providing the series.
3 In the case that both hypotheses in Equations (Equation3(3) ) and (Equation4
(4) ) are rejected, then we have the case of bidirectional causality. Bidirectional causality between military spending and economic growth implies a feedback system where both variables react to each other. If the hypothesis in Equation (Equation3
(3) ) is rejected, then military spending Granger causes economic growth. Analogously, if the hypothesis in Equation (Equation4
(4) ) is rejected, economic growth Granger causes military spending. It is also possible to have a case of no Granger causality in either direction implying that neither of the two variables have predictive content for each other.
4 See the Appendix of Balcilar and Ozdemir (Citation2013) for technical details of the bootstrap procedure.
5 Indeed, Granger (Citation1996) argued that parameter non-constancy was one of the most challenging issues confronting empirical studies.
6 Details of the rolling window technique are also explained in the Appendix of Balcilar and Ozdemir (Citation2013).
7 The critical values are reported in Andrews (Citation1993) and Andrews and Ploberger (Citation1994).
8 To consider the robustness of the tests to structural change, the Gregory-Hansen was used finding no cointegration and confirming the long-run instability result. The Zivot-Andrews tests also gave the same results as our unit roots when structural changes are incorporated. That is series are I(1), even when we allow for structural breaks in the individual series.
9 Lc test is a general fluctuation process-based test for parameter constancy. Under the alternative parameters vary according to a random walk process. It is based on the scores and can be used to check stability of all the parameters or subsets of the parameters. If the test is based on only the score of the intercept in a levels regression, then it can be interpreted as a cointegration test, because a random walk intercept implies a unit root. Here it is not used it to check for cointegration, but to check the stability of the parameters of the VAR model, so it cannot be interpreted as the cointegration test. If we had reported the long-run equation stability tests separately for each parameter than it can be used as cointegration tests. We used Lc in the sense of parameter constancy test based on the scores of the VAR model, not as the cointegration test, which would based on the score of the intercept in the levels long-run regression.
10 Both the Ave-F and the Exp-F statistics test the overall constancy of the parameters and are optimal tests as shown by Andrews and Ploberger (Citation1994).
11 This excludes the observations required for lags and hence is the actual number of observations in the VAR.