3.1. A two-regimes logit EWS for debt crisis
As said before, in this first specification we allow the independent variables to have a different marginal effect on the probability of ‘entering’and of ‘being’ into a debt crisis:3 more formally, entering into a crisis refers to the first year of the episode, while being into a crisis refers to the remaining years until the end of the negative event.
The dataset used in the logit regression includes the 28 macroeconomic variables that have been the subject of the event study analysis: the entire sample runs from 1980 to 2002. This means that we used macro data from 1980 to 2001 in order to explain the occurrence of debt crisis events from 1981 to 2002: therefore, the final regression sample contains something as 616 observations.4
Our estimation approach foresees three different steps:
– first of all, we run logit regressions for each of the 28 variables independently from one another and we exclude all the variables that turn out to be insignificant in determining the ‘entering’ or the ‘being’into a debt crisis, as well as the variables that, although being significant, have a counterintuitive sign;
– in a second phase, we run group-wise regressions, i.e. we group in families, essentially according to their nature, all the variables that got through the first step and then we run new logit regressions for each of these groups: as in the first phase, we retained only those variables that turned out to be significant and had the correct sign;
– in the final step, we put together all the variables that got through the first and the second step into a general logit regression. We added back variables that, for example, have been found to be significant in the literature or that displayed a particular behaviour in the event study analysis, but were dropped in either the first or the second step:5 using a ‘general to specific’approach in order to reach a parsimonious model, we again drop all the insignificant variables.At the end of this procedure, we retain just six variables: interest payments on external debt scaled to international reserves; the degree of openness to international trade; the export growth rate; the ratios of total external debt and short term debt to GDP; the ratio of international reserves to total external debt. Table 2 reports the estimation coefficients.
Testing whether the marginal effects on the probability of entering into a crisis or being into a crisis are the same led us to reach the final specification, which is reported in Table 3.
The probability of entering, or being, into a debt crisis seems to depend essentially on the behaviour of variables that measure the burden of external indebtedness and the foreign-currency generating capabilities of a country: