4.4. Análisis de variables e indicadores de la investigación
4.4.4. Cumplimiento de leyes aduaneras para facilitar el comercio
request).
application of the IMF m ethodology w o uld be conducive to assigning the EVI a
central role in guid in g the volum e and grant allocation of IDA aid, in the sam e
Graph 3.2: Sustainable Debt Level and EVI
(At 25 Percent Probability of Distress)
25 50 75
EVI Distribution (percentiles)
s Debt-to-GDP ■0 Debt-to-XGS — * — Debt-to-GDR
Data Sources: Author's calculations, b ased on UN, World Bank and IMF data. Debt ratios from the d ataset m ade available b e Art Kraay.
m anner an d w ith the sam e degree of confidence as the IMF analysis attem pts to
establish the prim acy of the CPIA. The policy im plications from this assertion should
be clear. For exam ple, w ith aid and g rant allocation centred a ro u n d the EVI, rather
than CPIA, countries w ith low vulnerability (25th percentile) w ou ld be deem ed
capable of carrying alm ost twice the volum e of debt as low -vulnerability countries
(75th percentile), w hile facing the sam e probability of debt distress consistent w ith the
debt-to-exports ratio. As a result, the relative share of aid allocation w ould tilt
tow ards low -vulnerability countries, w hile the relative w eigh t of grant financing
w o uld increase in favour of highly vulnerable countries. W hile it should be
em phasised th a t w e w o u ld no t necessarily argue in favour of centring aid allocation
on the EVI in such a m echanical fashion and w ith the im plications described, o ur key
point of contention is th at the IMF em pirics is geared to w ard s an essentially
un su b stan tiated establishm ent of the CPIA as a gu iding principle of debt
sustainability, at the expense of vulnerability concerns.
Next, we test the statistical significance of economic vulnerability indicators in
conjunction w ith the central policy indicators deployed by the IMF, fitting a num ber
of alternative pro bit m odels to the sam e dataset underly ing the p revious regressions.
A d opting d ebt distress episodes (identified, as before, according to the IMF
definition) as the b inary dep en d en t variable, the perform ance of each predictor is
first tested in conjunction w ith the debt-to-exports ratio an d a constant term. A nnex
Table A3.4 sum m arises the outcom e of the prelim inary estim ations, w hile Table A3.5
lists pair-w ise correlations am ong the predictors. It should be no ted th at w e do not
rep ort the results relating to regressions involving the debt-to-G D P ratio, instead of
the exports ratio, since they are qualitatively identical across all the specifications
envisaged. M oreover, w e exclude the debt-to-revenue ratio from o ur analysis, due to
the above m entioned d ata lim itations relating to dom estic g ov ernm en t revenue series
in the available databases. A gain, all explanatory variables used for regressions are
described in A nnex Table A3.1. W ith the exception of the aid flow s and the term s of
trade indexes, all predictors are lagged by one period to control for endogeneity.
W ith reg ard to the single prob it estim ations listed in Table A3.4, w e find th at all the
indicators are significant and enter the regressions w ith the expected sign. T hat is,
countries associated w ith h igher debt ratios; a relatively hig h er (i.e. w orse) CPIA
ranking; a low er (i.e. w orse) governance or rule-of-law ranking; a higher (i.e. w orse)
econom ic vulnerability ranking (as m easured by either the EVI, instability, or
concentration indices, or by a com bination of the instability and concentration
indices); and those w ith relatively low er aid inflows are associated w ith a higher
likelihood of experiencing deb t distress over time. W ith regard to the GT Index,
m easurin g the term s of trade deviation from four-year m oving averages, w e note a
relatively low er significance of its estim ated coefficient (at the five p er cent level),
taking a positive sign.38 If confirm ed by the m ore elaborate m odel specifications
below, w e are inclined to in terp ret this indicator's positive sign as an indication of
the distress-inducing effect of relatively low er term s of trad e in the years preceding
the distress episodes, considering th a t it enters the regressions w ith no tim e lag. As a
result, the relative term s of trad e increase occurs at a tim e w h en distress is already
38 Also the pair-wise correlation of the GT Index with the remaining predictors is far lower (see Annex Table A3.5).
manifest, and the indicator m arks a situation of im provem ent on the previous years
of relatively depressed price conditions.
Table 3.5 sum m arises the results from three further series of probit estim ations,
testing the perform ance of CPIA an d alternative m easures of policy perform ance
against the various vulnerability indices. The first group of estim ations (labelled
Probit 1 to Probit 5) assesses the significance of the CPIA after including each of the
vulnerability indicators taken individually. The outcom e is striking: the CPIA ceases
to be significant w h en tested in conjunction w ith the EVI, its instability and
concentration com ponents, or the joint instability and concentration term. Instead,
each of the econom ic vulnerability indicators enters the regression w ith the expected
sign and w ith the highest degree of significance (i.e. w ith a p-value below the 0.1 per
cent threshold). As an exception, the CPIA is show n to m aintain a low degree of
statistical significance in com bination w ith the GT Index, w hich is how ever explained
by the generally low perform ance of the GT Index if used in isolation from the m ore
com prehensive vulnerability indices (see below). Finally, the debt-to-exports ratio
confirms, as expected, its hig h degree of significance in pred icting debt distress
episodes.
In ord er to test these central results m ore thoroughly for statistical robustness, we
apply the likelihood-ratio (LR) test in addition to the stan d ard W ald test, the results
of w hich are associated w ith the nu m ber of stars next to the estim ated coefficients in
Table 3.5. In the case of m axim um -likelihood (ML) estim ations, the W ald test
involves assessing the null hypothesis w ith regard to any of the coefficients' taking
value zero, along a stan d ard tw o-tailed test on the basis of the z-statistic.39 In contrast,
the LR test assesses the sam e hypothesis by com paring the log-likelihood from the
full m odel w ith those of a restricted m odel im posing certain constraints. For example,
in order to test for the significance of the CPIA and EVI in the case of the Probit 1
m odel in Table 3.5, w e first com pute the log-likelihood of the full m odel, including
the debt ratio, CPIA, EVI and a constant term . We then re-estim ate the log-likelihood
39 Essentially, the Wald test for ML estimations mirrors the Wald test for ordinary least squares