complications (Sarte, 1999). While both short- and long-term restrictions are sensitive to
the exact parameter values imposed, substantially more uncertainty surrounds the
estimates of the long-term inverted moving average representation in (5.2), especially in
the short samples that we use (Christiano et al., 2006). The basic problem is that no
asymptotically correct confidence intervals on C(1) can be constructed. Faust and Leeper
(1997) prove that there are no consistent tests for the significance of the long-term
response. Specifying a priori the lag length of the VAR or choosing the horizon at which
the long run effect nullifies can solve this problem. One may also check the consistency
of some short-term restrictions with the long-term behaviour of the model, as in King and
Watson (1997). Third, there is a possibly large set of underlying shocks from which we
extract only a few. As discussed above, we extract a generic supply and cyclical shock, as
well as a fiscal shock. This necessarily involves a debatable linear aggregation over
shocks. If each shock affects the economy in qualitatively the same way the shocks may
be commingled. This is particularly acute for the analysis of fiscal policy, as different
expenditure and revenue categories may indeed have different longer run effects on
output that are not distinguishable from technology shocks but moreover have similar
short-term responses. Fourth, a problem may also occur of high frequency feedbacks. We
observe fiscal policy only at an annual frequency. We assume the structural shocks to be
orthogonal but if there are mid-year revisions of the budget, this may muddle both
economic and fiscal shocks. This only stresses the problem of correctly identifying the
timing of shifts in fiscal policies. Finally, a major assumption underlying the VAR-model
is parameter constancy. The conclusions of VARs are highly sensitive to the presence of
structural breaks. Especially for fiscal policy, there is evidence of non-linear effects (see
Giavazzi et al., 2000, for instance). We therefore run some stability tests on the VAR-
model.
5.4 Empirical Analysis
Data
All data are annual and come from AMECO.87 This database covers the longest available
period since 1970 up till 2004 for which fiscal data are available for France, Germany,
Portugal and Spain. Fiscal data and output are deflated by the GDP-deflator and are
defined in first differences of log-levels. In many studies, the fiscal data are scaled to
GDP, but this clouds inference. As economic shocks affect both fiscal variables and GDP,
this leads to a spurious negative correlation between the deficit and these shocks.
Moreover, we are primarily interested in distilling a fiscal indicator on the basis of the
historical decomposition of output. For the same reason, we do not concentrate on the
effects of fiscal policy on private output but use total output instead. We also ignore
possible cointegration between overall expenditures and revenues, which derives from the
intertemporal budget constraint.88 This implies that parameter estimates may no longer be
efficient albeit still consistent. However, inference on the short-term results of the VAR
would hardly be affected by non-stationarity of the data (Sims et al., 1990).
87
88
Details are in Appendix 5.A. A program containing the RATS-code for the SVAR model is available
from the authors upon request.
For such an analysis, see Claeys (2004).
134