MOHAMMAD S HASAN TABLE 1 Unit Root Test Results A. Stationarity Variable First difference PM 0.162(-3.52) 4.16(-2.93) -473(-3.52) 998(-2.9 0.395(-3.52 5.13(-2.93) 527(-3.52) 2.88(-2.93) 0.940(-3.52) 442(-2.93) -4.32(-3.53) -2.53(-2.93) 247(-3.52 -5.15(-2.93) 524(-3.53) 925(-2.93) 2.82(-3.52) 3.83(-2.93) 461(-3.53) 0.458(-2.93) -6.46(-2.93) 6.41(-3.52) I, and I- are the t-statistics based on augmented Dickey-Fuller(ADF)regression with allowance for a constant and trend, respectively. Figures in parentheses are McKinnons(1991) critical value equations reject the null hypothesis of noncointegration. The estimated coeffi cients are plausible in static multivariate regressions. The results suggest that, in most cases, variables are statistically significant and have the signs predicted by the theory Given the fact that a residual-based testing procedure, such as Engle-Granger ( 1987), is alleged to have a low power to detect an otherwise dormant long-run relationship, we proceed to test for the presence of cointegrating vectors using the Johansen and Juselius (JJ)maximum likelihood procedure. For further details of the JJ method, see Johansen(1988), Johansen and Juselius(1990). Since annual data loyed, a maximum lag length of two is used in the Johansen vector autoregressive(VAR) model. Results of Johansen's maximum eigenvalue and trace tests are presented in Table 3. Cheung and Lai(1993) have argued that Johansen's likelihood ratio(Lr)tests are derived from asymptotic results and standard inferences in finite samples may not be appropriate Johansens LR tests re biased toward finding cointegration too often in finite samples when asymp- totic critical values are used. The finite sample bias of Johansens test is a positive function of T/(T-nk), where T, n, and k signify the sample size, the number of variables in the estimated system and the lag length, respectively Reimers(1991), and Reinsel and Ahn(1992) have suggested adjusting Jo- hansens test statistics by a scaling factor of (T-nk)/T and comparing them with their asymptotic critical values. Following Reinsel and Ahn(1992), th computed test statistics were adjusted using the scaling factor. We have used the critical values furnished by Charemza and Deadman (1992). Following Cheung and Lai(1993), the critical values were also adjusted by multiplying by the degree-of-freedom correction factor, T/(T-nk), and then comparing with the asymptotic test statistics The results turn out to be similarequations reject the null hypothesis of noncointegration.8 The estimated coeffi- cients are plausible in static multivariate regressions. The results suggest that, in most cases, variables are statistically significant and have the signs predicted by the theory. Given the fact that a residual-based testing procedure, such as Engle–Granger (1987), is alleged to have a low power to detect an otherwise dormant long-run relationship, we proceed to test for the presence of cointegrating vectors using the Johansen and Juselius (JJ) maximum likelihood procedure. For further details of the JJ method, see Johansen (1988), Johansen and Juselius (1990). Since annual data are employed, a maximum lag length of two is used in the Johansen vector autoregressive (VAR) model. Results of Johansen’s maximum eigenvalue and trace tests are presented in Table 3. Cheung and Lai (1993) have argued that Johansen’s likelihood ratio (LR) tests are derived from asymptotic results and standard inferences in finite samples may not be appropriate. Johansen’s LR tests are biased toward finding cointegration too often in finite samples when asymp￾totic critical values are used. The finite sample bias of Johansen’s test is a positive function of T/(T 2 nk), where T, n, and k signify the sample size, the number of variables in the estimated system and the lag length, respectively. Reimers (1991), and Reinsel and Ahn (1992) have suggested adjusting Jo￾hansen’s test statistics by a scaling factor of (T 2 nk)/T and comparing them with their asymptotic critical values. Following Reinsel and Ahn (1992), the computed test statistics were adjusted using the scaling factor.9 8 We have used the critical values furnished by Charemza and Deadman (1992). 9 Following Cheung and Lai (1993), the critical values were also adjusted by multiplying by the degree-of-freedom correction factor, T/(T-nk), and then comparing with the asymptotic test statistics. The results turn out to be similar. TABLE 1 Unit Root Test Results A. Stationarity testa Variable Level First difference tm tp tm tp P 1.79 (22.93) 0.162 (23.52) 24.16 (22.93) 24.73 (23.52) M 0.998 (22.93) 20.395 (23.52) 25.13 (22.93) 25.27 (23.52) W 2.88 (22.93) 0.940 (23.52) 24.42 (22.93) 24.32 (23.53) g 22.53 (22.93) 22.47 (23.52) 25.15 (22.93) 25.24 (23.53) AP 0.925 (22.93) 22.82 (23.52) 23.83 (22.93) 24.61 (23.53) IP 0.458 (22.93) 21.15 (23.52) 26.46 (22.93) 26.41 (23.52) a tm and t p are the t-statistics based on augmented Dickey–Fuller (ADF) regression with allowance for a constant and trend, respectively. Figures in parentheses are McKinnon’s (1991) critical value. 676 MOHAMMAD S. HASAN