Cointegration and error correction models

G.E. Nasr et al. r Energy Economics 22 2000 627]640 635 The partial adjustment model explains 89 of the variation in electricity demand. The null hypothesis of no serial correlation cannot be rejected given the Ž . value of the Durbin h test. The DF and ADF 1 tests on the residuals indicate Ž . stationarity and integration of zero order, I 0 . All coefficients are significant at the 5 level except for DD. The coefficients of TI and DD are 0.15 and 0.11, respectively. The results indicate a major improvement over the 1993]1994 period. Once the Ž . autocorrelation of the residuals were accounted for, in the AR 1 model, all coefficients become significant at the 5 significance level with the right expected sign. The year 1995 represents a transition period for electricity production. Rationing hours which attained 12 h a day in 1993]1994 were significantly decreased during this year to 6.5 h a day. The results agree with many studies Ž found in the literature indicating the significance and impact of GPD Dincer and . Ž . Dost, 1997 and DD Al-Zayer and Al-Ibrahim, 1996 though considered sepa- rately, on electricity consumption. Specifically, the results indicate that an increase in GDP and DD implicates an increase in electrical energy consumption. 4.3. The 1996]1997 period In spite of the rationing hours that existed in 1995, analysis results of the 1995]1997 time period are found to be satisfactory. However, separate analysis for the 1996]1997 time period is performed since rationing was virtually non-existent for that period. Table 3 shows regression analysis results at the 5 significance level. The DW statistic for the static model indicates serial correlation of the residuals. The model R 2 is 0.13 and all determinants are insignificant at the 5 level. TI and DD coefficients are determined to be 0.22 and 0.39, respectively. The first order autoregressive model is also utilized to correct for serial correlation of the residuals. The model explains 54 of the variation in electricity demand. Again, all coefficients are significant at the 5 level including the first order autoregressive term, r. Furthermore, DD exhibits positive correlation to consump- tion. The performance of the partial adjustment model for this period appear to be as Ž . good as the AR 1 model. All coefficients are found to be significant at the 5 level except for DD. The model explains 46 of the variation in electricity demand. The null hypothesis of no serial correlation is not rejected using the Ž . Durbin h test. Again, the DF and ADF 1 tests indicate stationarity and integration Ž . of zero order, I 0 , of the residuals. TI and DD coefficients are 0.25 and 0.27, respectively. As rationing hours were virtually non-existent, starting 1996, in most of the country, the electrical energy imbalance of consumption and production was eliminated and electricity consumption started to be demand-driven.

5. Cointegration and error correction models

This section reports the results of the DF and ADF tests for unit roots in the G.E. Nasr et al. r Energy Economics 22 2000 627]640 636 Table 3 a Regression analysis results for the 1996]1997 time span Ž . Coefficient Static model AR 1 model Partial adjustment model A 411.2 479.4 y 7.0 Ž . Ž . Ž . 3.27 6.15 y 0.046 Ž . B TI 0.22 0.17 0.25 1 Ž . Ž . Ž . 1.80 2.41 2.47 Ž . B DD 0.39 0.5 0.27 2 Ž . Ž . Ž . 1.55 2.37 1.25 Ž . B C ] ] 0.61 3 ty 1 Ž . 3.69 r ] 0.74 ] Ž . 4.31 2 R 0.13 0.54 0.46 DW or h 0.89 2.36 0.55 DF y 3.32 y 5.50 y 4.18 Ž . ADF 1 y 3.62 y 3.14 y 2.52 a Figures in parentheses are t-statistics. variables for the periods 1995]1997 and 1996]1997 and the subsequent analysis of cointegration and estimation of Error Correction Models. The results in the levels and differences for the variables C, TI and DD are reported in Table 4. The results show that we could not reject the null hypothesis of unit roots for the level series. However, the null hypothesis was rejected for the first differences, indicat- Ž . ing that the variables are first-difference stationary, I 1 , for both periods. Given these results, the Johansen, and Engle and Yoo tests are used to check for cointegration. The results of the Johansen cointegration test, which is based on a Ž . maximum likelihood vector autoregressive VAR procedure, are shown in Table 5. Table 4 DF and ADF test for stationarity Ž . DF ADF 3 Levels 1st difference Levels 1st difference 1995]1997 period a a C y 2.652 y 7.215 y 2.059 y 5.361 a b TI y 3.187 y 9.108 y 2.901 y 4.039 a b DD y 3.409 y 3.845 y 2.625 y 3.216 1996]1997 period a a C y 3.317 y 6.046 y 2.809 y 5.220 b a b TI y 5.473 y 9.662 y 2.258 y 3.690 b a DD y 2.781 y 3.01 y 3.517 y 4.997 a Significant at 1 level. b Significant at 5 level. G.E. Nasr et al. r Energy Economics 22 2000 627]640 637 Table 5 Ž . Johansen cointegration test VAR 1 test of cointegration vector Eigenvalue Likelihood 5 critical 1 critical Hypothesized Ž . ratio value value no. of CE s 1995]1997 a c 0.404224 30.96722 29.68 35.65 None a 0.257806 12.84107 15.41 20.04 At most 1 a 0.066434 2.406032 3.76 6.65 At most 2 1996]1997 b d 0.825823 55.64383 29.68 35.65 None b c 0.511556 17.19485 15.41 20.04 At most 1 b 0.062982 1.431164 3.76 6.65 At most 2 a Ž . L.R. test indicates 1 cointegrating equation s at 5 significance level. b Ž . L.R. test indicated 2 cointegrating equation s at 5 significance level. c Denotes rejection of the hypothesis at 5 significance level. d Denotes rejection of the hypothesis at 1 significance level. The trace statistic, L.R. test, indicates the presence of at most one and two cointegrating equations for 1995]1997 and 1996]1997 subsets, respectively, at the Ž . 5 significance level. These results are corroborated by the Engle and Yoo 1987 Ž . procedure for cointegration. The residuals from Eq. 4 for both subsets are tested using DF and ADF tests. Both tests indicate stationary residuals and integration of Ž . zero order, I 0 , as shown in Tables 2 and 3. The error correction models are estimated. The coefficients of the lagged error term are found to be y0.18 and y 0.23 for both subsets, respectively, and are significantly different from zero. The Table 6 Error correction model of electricity consumption Coefficient 1995]1997 period 1996]1997 period Parameter estimate Parameter estimate D 10.69 5.94 o Ž . Ž . 1.48 0.661 Ž . D DC y 0.25 y 0.30 1 ty 1 Ž . Ž . y 1.47 y 1.48 Ž . D DTI 0.10 0.11 2 Ž . Ž . 1.64 1.64 Ž . D D DD 0.44 0.623 3 Ž . Ž . 2.82 3.20 Ž . D u y 0.18 y 0.23 4 ty 1 Ž . Ž . y 2.5 y 1.34 2 R 0.38 0.49 DW 1.90 1.93 DF y 5.45 y 4.75 Ž . ADF 1 y 3.71 y 4.28 G.E. Nasr et al. r Energy Economics 22 2000 627]640 638 residuals are stationary and not correlated. The results for the ECM models are shown in Table 6.

6. Forecasting error analysis