Regarding the econometric approach, there are two major strands in the literature. The
first uses linear regression methods on a numerical representation of the ratings. The
early study by Cantor and Packer (1996), applies OLS regressions to a linear
representation of the ratings, on a cross section of 45 countries. This methodology was
also pursued by Afonso (2003), Alexe et al. (2003) and Butler and Fauver (2006). Using
OLS analysis on a numerical representation of the ratings is quite simple and allows for
a straightforward generalization to panel data by doing fixed or random effects
estimation (Monfort and Mulder, 2000; Eliasson, 2002 and Canuto et al., 2004).
Although estimating the determinants of ratings using these approaches has in general a
good fit and a good predictive power it faces some critiques. As ratings are a qualitative
ordinal measure, using traditional estimation techniques on a linear representation of the
ratings is not the most adequate framework of estimation. First, it implies the
assumption that the difference between two rating categories is equal for any two
adjacent categories, which would need to be tested. Furthermore, even if this
assumption was true, because of the presence of elements in the top and bottom
category, the estimates are biased, even in big samples. Nevertheless, Eliasson (2002)
argues that given the existence of many categories one can treat the rating variable as
continuous and to overcome the criticism of the assumption of an even distance between
steps, it is possible to use different quantitative transformations. For instance, Reisen
and Maltzan (1999) apply a logistic transformation of the ratings and Afonso (2003)
applies both a logistic and an exponential transformation of the ratings. In that case, the
differences between categories are not constant, but are still imposed a priori.
The other strand of the literature uses ordered response models. Because the ratings are
a qualitative ordinal measure, the established wisdom advises the use of ordered probit
estimation. This method will itself determine the size of the differences between each
category. For example, this procedure was used by Hu et al. (2002), Bissoondoyal-
Bheenick (2005) and Bissoondoyal-Bheenick et al. (2005). Although this should be
considered the preferred estimation procedure it is not entirely satisfying. The crucial
point is that the ordered probit asymptotic properties do not generalise for a small
sample, so if we estimate the determinants of the ratings using a cross-section of
countries, we would have too few observations. It is therefore imperative to try to
maximize the number of observation by using panel data, but when doing so, one has to