Stability of money demand function revisited in China

Abstract

As China's economic reforms have undergone significant structural changes after 1979, it has been rather difficult to formulate a stable money demand function over the period following that year. While previous literature on the long-run relationship of money demand in China shows the existence of stable money demand, this article revisits the stability of the China money demand function over the period after 1979. To employ the unit root tests and the cointegration tests with structural break, the empirical evidence demonstrates that economic and financial deregulation did affect the stability of demand for money in China over the period 1977 to 2002. Moreover, the estimated long-run income and interest elasticity are respectively 1.01 (1.11) and −0.14 (−0.08) using the real M1 (M2) equation. In addition, real income and the interest rate are found to be weakly exogenous. We overall do find structural breakpoints mainly in 1980 and 1993, and they look to match clearly with corresponding critical financial and economic incidents.

Acknowledgements

This work is supported by the Ministry of Education, Taiwan, R.O.C. under the ATU plan.

Notes

¹See Siriam (1999).

²Chaisrisawatsuk et al . (2004) indicate that the relative absence of empirical money demand studies for transition economies is due in part to the relative instability of these economies in the transition process itself as well as because of concerns over the reliability and frequency of time series data.

³See Xu (1998), Yu and Tsui (2000) and Qin et al . (2005) for a detailed discussion.

⁴See Yu and Xie (1999), p. 34.

⁵Beforehand, the reforms of macroeconomic policy were determined solely by the State Council in China. The central bank did not have independent monetary policies, prompting it to essentially support the implementation of the physical output targets contained in the central plan. See Blejer et al . (1991).

⁶They capture the effects of economic and financial reforms subjectively by including a dummy variable that takes a value of one for the period 1979 to 1988 and zero elsewhere.

⁷Traditional breakpoint tests include the Wald test Likelihood ratio test, Lagrange multiplier test and Chow test, etc.

⁸Pradhan and Subramanian (2003) and Ramachandran (2004) offer detailed discussions.

⁹Our justification for using annual data is based on the following. As Davidson and MacKinnon (1993) state, in order to avoid any potential biases from using seasonally adjusted data (quarterly or monthly data) when conducting unit root tests, annual data should be used. Along the same lines, Hakkio and Rush (1991) find that when using monthly or quarterly data in a cointegration analysis, increasing the number of observations does not add any robustness to the results.

¹⁰Nominal income measures gross output from agriculture, industry, construction, transportation, and commerce.

¹¹Poole (1988) argues that long-term interest rate specifications are more robust than those employing a short-term interest rate in the money demand function.

¹²See Ericsson et al . (1998), Carlson et al . (2000) and Cargill and Parker (2004).

¹³For the low power of finite-sample problems, ZA (1992) propose some explanations in Section V. In ZA, the authors use the Nelson–Plosser data set to carry on endogenous break unit root tests. We find that the observations (T) of some series (real GNP, nominal GNP, real per capita GNP) are only 62 (same as Table 3 of Perron, 1997), but ZA adopt asymptotic critical values equally (see p. 259, Table 6), in which critical values for the limiting distributions in the theorem are obtained by simulation methods. We follow ZA's approach and adopt asymptotic critical values provided by ZA. Follow-up related empirical papers, like Cakan and Özmen (2002) and Kollias et al . (2004), also apply the asymptotic critical values as a guideline for their empirical results when they use the endogenous break unit root tests.

¹⁴See Huang (1998, p. 6).

¹⁵See Mchran and Quintyn (1996, p. 25).

¹⁶See DaCosta and Foo (2002, p. 4).

¹⁷Bessler and Binkley (1982) and Geweke and Meese (1981) show that the SBC appears to be superior to other lag length selection methods.

¹⁸Because all of the variables used in this study have unit-root characteristics, using nonlinear models such as the threshold autoregressive model or the smooth transition autoregressive model to deal with the problem of structural change among the variables is also not appropriate. This is because those models’ starting points require that all variables conform to the stationarity premise. For more details, refer to Terasvirta (1994).