Most hypothesised structural breaks appear after 2000, but none of the t-statistics associated
with the possible structural break dates are statistically significant at .95% confidence level.
Thus the results of the Zivot-Andrews test reinforces the DF-GLS test result, thus we consider
all series being integrated of order 1.
4.2. Cointegration tests
First, the VECM lag length was selected. The various lag length criteria suggested different
lag lengths, ranging between 1 (Schwarz-Bayesian Criterion), and 12 (Akaike Information
Criterion). 5 lags in the VAR model were considered enough to result uncorrelated residuals,
the Final Prediction Error and LR statistic also selecting the same lag length. The Pantula
principle selected model 4, where there is a trend restricted to the cointegration space. The
cointegration test results are presented in table 4 and 5.
Table 4. Johansen cointegration test results - trace statistics
|
Hypothesized |
Eigenvalue |
Trace Statistic |
5% Critical Value |
1% Critical Value |
|
None |
0.326620 |
90.97482 |
62.99 |
70.05 |
|
At most 1 |
0.279446 |
56.96656 |
42.44 |
48.45 |
|
At most 2 |
0.220366 |
28.78137 |
25.32 |
30.45 |
|
At most 3 |
0.082164 |
7.373335 |
12.25 |
16.26 |
Table 5. Johansen cointegration test results - max Eigen statistics
|
Hypothesized |
Eigenvalue |
Max-Eigen Statistic |
5% Critical Value |
1% Critical Value |
|
None |
0.326620 |
34.00826 |
31.46 |
36.65 |
|
At most 1 |
0.279446 |
28.18519 |
25.54 |
30.34 |
|
At most 2 |
0.220366 |
21.40804 |
18.96 |
23.65 |
|
At most 3 |
0.082164 |
7.373335 |
12.25 |
16.26 |
The trace statistics selects 3 cointegration vectors at 5% level and 2 cointegration vectors at
1%, level, whilst the maximum Eigen statistic selects 3 cointegration equations at 5% level.
11