In a similar study, Demetriades and Hussein (1996) conduct causality tests between financial development and economic growth using cointegration and Granger causality techniques for 16 countries. Their results provide little support for the notion that financial development is a leading factor in the process of economic development.
     They find considerable evidence of bidirectionality and some evidence of reverse causation. More recently, Ghali (1999) investigates whether financial development leads to economic growth in the small developing economy of Tunisia. His results suggest the existence of a stable long-run relationship between the development of financial sector and the evolution of per capita real output that is consistent with the view that financial development can be an engine of growth in this country. Using cointegration and Hsiao’s version of the Granger causality method, Cheng (1999) finds causality running from financial development to economic growth with feedback in post-war South Korea and Taiwan. These results support the Patrick (1966) hypothesis that there is likely to be an interaction of supply-leading and demand-following phenomena. Most of the previous studies focus only on a two-variable case and their results may be biased due to the omission of relevant variables. Recent empirical studies have addressed this shortcoming. For example, Luintel and Khan (1999) examine the long-run relationship between financial development and economic growth using multivariate VAR models for ten countries. They find that the long-run financial development and output relationships are identified and bidirectional causality between financial development and economic growth exists for all sample countries. On the other hand, Darrat (1999) uses multivariate Granger causality tests within an error-correction framework to investigate the role of financial development in            economic growth in three middle-eastern countries, namely, Saudi Arabia, Turkey, and the United Arab Emirates, and his results generally support the view that financial development is a necessary causal factor of economic growth. Although much of the recent evidence seems to indicate that financial development causes economic growth, the issue for Taiwan is unresolved. In this paper these new time series methods are used to investigate the relationship between financial development and economic growth in Taiwan. III. Data The empirical analysis is based on annual data on real GDP per capita, M2, exports, and imports for Taiwan over the period of 1962 to 1998 (1991 . 100). Following most of the literature (Jung, 1986; Cheng, 1999; Darrat, 1999), financial development is calculated as the ratio of M2 to GDP.1 All the data series are transformed into logarithms to achieve stationarity in variance. Data are obtained from the AREMOS database of the Taiwan Ministry of Education. 1 An alternative measure calculated as the ratio of liquid liability to GDP was also used in this study. Results are similar to those reported here and are available upon request from the authors. 

     Financial development and economic growth IV. Methodology and Empirical Results Unit root tests A number of authors have pointed out that the standard ADF test is not appropriate for variables that may have undergone structural changes.2 For example, Perron (1989, 1990), Banerjee et al. (1992), and Zivot and Andrews (1992) have shown that the existence of structural changes biases the standard ADF test towards non-rejection of the null of unit root. Hence, it might be incorrect to conclude that the variables are nonstationary on the basis of the results using the standard ADF tests.3 To address the problem, Perron (1990) developed a procedure for testing the hypothesis that a given series fYtg has a unit root, given that an exogenous structural break occurs at time TB. Zivot and Andrews (1992, hereafter ZA) criticized this assumption of an exogenous break point and developed a unit-root test procedure that allows an estimated break in the trend function under the alternative hypothesis. Therefore, it seems appropriate to treat the structural break as endogenous and test the order of integration by the ZA procedure. The ZA tests are represented by the following augmented regression equations: 1 t t  AModel A:  Yt .  A t  A 2 DUt1 k Xt  AYt 1 t  j  Yt j t " t j.1 equation for every possible break point within the sample and calculates the t-statistic for the estimated coefficients. This tests the null hypothesis of a unit root against the alternative hypothesis of trend stationarity with a one-time break (TB) in the intercept and slope of the trend function at unknown point in time. The null of a unit root is rejected if the coefficient of Yt . 1 is significantly different from zero. The selected break point for each data series is that value of TB for which the t-statistic for the null is minimized. Since the choice of lag length k may affect the test results, the lag length is selected according to the procedure suggested by Perron (1989). This procedure involves starting with an upper bound kmax for k. If the last included lag is significant, then choose k . kmax. If not, reduce k by 1 until the last lag becomes significant.