Estimation results obtained with various econometric methods are presented in Table 3.
Fixed effect estimates confirm the probit regression analysis, with the exception of the
unemployment rate variable which is found to be insignificant (column 1).100 In particular, both
the level of the cyclically-adjusted fiscal position and the change in the cyclically-adjusted
primary fiscal balance are significant with expected signs. The country fixed effects are also
found to be significant, pointing to cross-country differences in policy preferences and/or the
influence of omitted variables on policy stances.101 However, fixed effects estimates can not be
relied on since, as already noted, many of the explanatory variables -including the fiscal ones- are
likely to be endogenous and thus to bias the results.102
Endogeneity problems are tackled in column 2 by means of system-GMM estimates
(Blundell and Bond, 1998). This method is particularly appropriate for the estimation of dynamic
panel data equations when the dependent variable is highly persistent and/or certain explanatory
variables are endogenous (see e.g. Bond, 2002). Here, all explanatory variables are assumed to be
endogenous, and standard statistical tests suggest that the model is well specified.103 The main
difference between System-GMM and fixed effect estimates is that the lagged policy stance
becomes significant, suggesting that reforms are more likely to occur when the initial policy
stance is strict. However, this effect is quantitatively small, pointing to a high degree of inertia.
Both the level of the fiscal balance and the fiscal adjustment variable remain significant, although
only at the 10% as regards the latter.
100
The lack of significance of the unemployment rate can be explained on two grounds. First, its effect may already be captured - at least
partly - by the crisis variable. Second, within the fixed-effect framework adopted here, the unemployment rate is likely to work essentially as a
proxy for the business cycle position.
101
102
A Fisher test rejects the null hypothesis of no fixed effects at the 1% level.
Another potential concern is the standard downward dependent variable bias (Nickell, 1981). However, the downward lagged dependent
variable bias falls as the time span of the sample increases. Furthermore, it is less of a concern when the time span is large and of the same order
of magnitude of the number of countries, as is the case here (Judson and Owen, 1999).
103. The Hansen test of over-identifying restrictions does not reject the null hypothesis of valid moment conditions at the 5% (and even the
10%) level, and the Arellano-Bond test rejects the null hypothesis of no first-order autocorrelation in the residuals but accepts -as would be
expected- the null of second-order autocorrelation. However, the power of the Hansen test to detect invalid overidentifying restrictions can
decline dramatically in small samples if an excessive number of moment conditions is used (Bowsher, 2000). Therefore, the system-GMM
estimates presented in Table 3 only use instruments (levels of explanatory variables) lagged 2 and 3 periods in difference equations and
instruments (first differences of explanatory variables) lagged 1 period in level equations.
197