convergence among the Northern regions. The same test on the y reaches
exactly the opposite conclusion of no convergence.
In order to investigate whether there is heterogeneity in the coefficients
of our regressions, we have used the Hausman statistic discussed in section
3 to perform a test of poolability of the data. Table 4 shows that according
to this test the slopes of the autoregressive parameters are homogeneous at
any level of aggregation. This result is not surprising since our measures
of TFP are bound to capture the more persistent part of growth dynamics.
Hence this test simply concludes that the mean-reversion (or non-reversion)
properties of each series are very similar.12
As a second step, we have used the test proposed by Im, Pesaran and
Shin (IPS) in order to take into account not only of the fixed effects hetero-
geneity (as in LLC) , but also the possible presence of heterogeneity in the
autoregressive parameter, a restriction that may be particularly binding in
our case. Applied to the entire panel, this test cannot reject the null of no
stationarity, using both measures of distance. This result basically mirrors
the one obtained with the LLC test. However, this test concludes for no con-
vergence even when we break down the sample into sub-groups of regions.
Only in one instance, the test is able to reject the null, namely among the
regions of the South and when we use y.
IPS, however, does not consider fully the potential cross-sectional inter-
dependence between the regions. A solution to this problem is the use of
the SURE methodology. Therefore, we have performed the test proposed
by Sarno and Taylor (1998) and Taylor and Sarno (1998), MADF. In this
test, rejection of the null cannot be taken as conclusive evidence that all the
series are stationary, but that at least one is. The interesting result, when
we apply this test, is that the null hypothesis is often rejected in the case of
y, leading us to conclude that some regions must be in a convergence process
12 A corollary of this test could be that pooling the data as in the LLC test is not
particularly damaging. It is however, important to note that the test may be affected by
inconsistencies deriving from the dynamic nature of the data
14