regressors are correlated with the residuals and the corresponding estimators also have a
normal limit distribution.
Using the software recently developed by Chiang and Kao (2001) we estimated our
equations using the POLS, PCOLS, DPOLS and the PFMOLS estimators. The results
obtained by the first three estimators are basically similar. In such regressions most
coefficients appear non-significantly different from zero or wrong signed. In contrast the
results supplied by the PFMOLS estimator are quite reasonable in terms of both sign and
magnitude. The small sample, the correlation in the residuals as well as the endogeneity of
some of the regressors probably explains these differences. For this reason, below we only
present and comment the PFMOLS results.
The reported equations below only consider two and not three interest rates as (7.1)
would suggest. The point is that due to strong colinearity it is very difficult to separately
estimate the coefficients associated with the interest rates. But the fact is that in this case
the exclusion of it from (7.1) is likely not to have damaging consequences for the
interpretation of the results of the estimated equations25.
As a matter of fact it turned out to be non-significant in the estimations. Therefore in
the regressions reported in Table 5, which were obtained using PFMOLS, it was excluded.
Below each coefficient is the computed t-statistic, which is asymptotically normal
distributed. For each equation several cointegration tests were computed (but are not
reported for space reasons). The null of a unit root in the residuals was always rejected, so
that all the equations presented in Table 5 are valid cointegrating relations26.
Column 1 displays the results of our basic specification (equation (7.1)) with α4 set
equal to zero). It can readily be seen that all the coefficients are statistically significant and
exhibit the expected sign for a loan-supply function. This, of course, is a strong piece of
evidence favouring our identification approach27.
25 If we, quite realistically, assume that it and st are cointegrated we may write it=st+k+εt where k
is a constant and εt a purely stochastic stationary process. The relevant part of the model may be written as
α3lt+α4it+α5st=α3lt+(α4+α5)st+α4k+α4εt, so that by not introducing it into the estimated model
we are subsuming the terms α4k and α4εt into the constant and the residuals of the resulting model
respectively, without significant consequences on the remaining estimated coefficients. However, the
coefficient on st should now be seen as being equal to (α4+ α5).
26 The panel cointegration tests computed by the NPT 1.2 package developed by Chiang and Kao (2001)
include the five panel cointegration tests developed in Kao (1999) and four panel cointegration tests developed
in Pedroni (1997).
27 The fact that all the estimated coefficients have the right sign does not, of course, completely rule out
the possibility of our estimated equation being a biased estimator of the true supply schedule. This will be the
26