be careful. Indeed, the generalization of ordered probit to panel data is not completely
straightforward, due to the existence of a country specific effect. Furthermore, within
this framework, the need to have many observations makes it harder to perform
robustness analysis by, for instance, partitioning the sample. In Section Three we will
address these questions when explaining our modelling strategy.
3. Methodology
Using a linear scale we grouped the ratings in 17 categories, by putting together in the
same bucket the few observations below B-. Indeed, if we used a specific number for
each existing rating notch, for instance 21 categories, it might be hard to efficiently
estimate the threshold points between CCC+ and CCC, CCC and CCC- and so on, given
that the bottom rating categories have very few observations. Table 1 above also shows
the relation established between the qualitative and the possible linear scales. Moreover,
and as we will see latter in the paper, a linear transformation is quite adherent to the
data. Nevertheless, we also report in Appendix 3 estimation results using a logistic
transformation.
3.1. Explanatory variables
Building on the evidence provided by the existing literature, we identify a set of main
macroeconomic and qualitative variables that may determine sovereign ratings.
GDPper capita - positive impact on rating: more developed economies are expected to
have more stable institutions to prevent government over-borrowing and to be less
vulnerable to exogenous shocks.
Real GDP growth - positive impact: higher real growth strengthens the government’s
ability to repay outstanding obligations.
Inflation - uncertain impact: on the one hand, it reduces the real stock of outstanding
government debt in domestic currency, leaving overall more resources for the coverage
of foreign debt obligations. On the other hand, it is symptomatic of problems at the
macroeconomic policy level, especially if caused by monetary financing of deficits.
10