regressors in levels. Accordingly, the t-ratios are adjusted to test hypotheses on the coefficients.
27
A second alternative is of course the Johansen (1988) procedure, in which a full vector
autoregressive system is estimated. However, in finite samples this procedure has the serious
problem of “the curse of dimensionality.” Monte Carlo evidence suggests that the performance
of this procedure is very poor in small samples. It generates frequent outliers and large mean bias
and the power of the tests are very low. More importantly, however, the procedure is less
effective than the single equation approach if the system parameters are misspecified (e.g. in
terms of lag length), and if there are problems like serial correlation in the equilibrium error
(Hargreaves, 1994, and Baffes, Elbadawi, and O’Connell, 1999). Thus, while in our case the
Johansen method is not suitable for estimation purposes, we do employ it to determine the
number of cointegrating vectors. To estimate the cointegrating vectors we employ DOLS. Then
the existence of cointegrating vectors is confirmed with relevant unit root tests of the residuals
obtained from the DOLS procedure.
In Tables 3 and 4, the test results from the Johansen procedure are reported. The trace
statistics for both Poland and Russia suggest the existence of one cointegrating vector at 1% as
well as 5% levels of significance. The maximum eigen value statistics suggest the existence of
one cointegrating vector for Poland and two cointegrating vectors for Russia at both 1% and 5%
levels of significance.
27 Descriptions of this procedure and the related adjustment methods can be found in Hamilton (1994, pp. 608 -
612) and Hayashi (2000, pp. 650 - 655).
21