we look at the fixed effects estimation we can see how poor it is. Notice the estimated
constant and its significance. In general the constant captures the middle rating, while
the estimated country dummies (ommited here), which vary from -7 to 7 notches,
capture each country’s average rating. All the other variables only capture small
movements from the rating in relation to its average4.
4.2.2. Differentiation across sub-periods
The separation of the overall sample into two sub-periods allows to assess broadly the
robustness of the empirical models and provides additional insight into possible changes
in the rating determinants. In particular, cutting the sample period in 2000 could reveal
any changes in the sovereign ratings methodology in response to the Asian crisis which
was perceived by market participants as revealing previously underestimated risks to
sovereign sustainability. Additionally, this also divides the full sample in two rather
similar sized sub-samples.
The models for the sub-periods are generally in line with those for the full estimation
period, although the significance levels of the individual coefficients are reduced (see
Tables 7-9). The lower significance levels reflect the reduction in the respective sample
sizes in half, which makes the coefficient estimates less certain. Taking this into
account, signs and orders of magnitude of the coefficient estimates from the full-period
models are mostly confirmed for the sub-periods. In particular, the core variables
identified above enter the models for the sub-periods with the correct sign and generally
significantly with a comparable order of magnitude.
[Insert Tables 7, 8 and 9 here]
Regarding the possible impact of the Asian crisis on ratings approaches, the stability of
the ratings models suggests that there was no fundamental change in methodologies. A
change that may point to some adaptation of ratings methods in response to the Asian
events is the decline in importance of the current account variable for Moody’s and
S&P, both with regard to the value of the estimated coefficient as well as its
4 We also estimated the model with the average rating of the three agencies, and also pooling the data for
the three agencies, but the results where quite similar.
19