Abstract
A large literature has used tests for Granger (1969) non‐causality, GNC, to examine the interaction of military spending with the economy. Such tests answer a specific although quite limited question: can one reject the null hypothesis that one variable does not help predict another? If one can reject, there is said to be Granger causality, GC. Although the limitations of GNC tests are well known, they are often not emphasised in the applied literature and so may be forgotten. This paper considers the econometric and methodological issues involved and illustrates them with data for the US and other countries. There are three main issues. First, the tests may not be informative about the substantive issue, the interaction of military expenditure and the economy, since Granger causality does not correspond to the usual notion of economic causality. To determine the relationship of the two notions of causality requires an identified structural model. Second, the tests are very sensitive to specification. GNC testing is usually done in the context of a vector autoregression, VAR, and the test results are sensitive to the variables and deterministic terms included in the VAR, lag length, sample or observation window used, treatment of integration and cointegration and level of significance. Statistical criteria may not be very informative about these choices. Third, since the parameters are not structural, the test results may not be stable over different time periods or different countries.
Notes
1 Micro‐econometrics uses a quite different concept of causality from GC, based on comparison of different potential outcomes: two hypothetical states of the world, one with the potential cause, usually labelled the treatment; and one without. Heckman (Citation2008) or Angrist and Pischke (Citation2009) provide discussions of this concept, and Lechner (Citation2006) discusses the relationship between them, but we confine ourselves to GC.
2 Granger originally defined it by subtracting information on the potential predictor from all possible information; but this is not operational, since one cannot specify all possible information.
3 There are other test procedures, e.g. using the cross‐spectrum, but these have not been widely used in defence economics – an exception is Gerace (Citation2002). These face similar difficulties to the VAR approach on which we concentrate.