Is the US demand for money unstable?

Abstract

The demand for money (M1) for the US is estimated with annual data from 1960 to 2008 and its stability is analysed with the extended Gregory and Hansen (1996b) test. In addition to estimating the canonical specification, alternative specifications are estimated which include a trend and additional variables to proxy the cost of holding money. Results with our extended specification showed that there has been a structural change in 1998 and the constraint that income elasticity is unity could not be rejected by subsample estimates. Short run dynamic adjustment equations are estimated with the lagged residuals from the Fully Modified Ordinary Least Squares (FMOLS) estimates of cointegrating equation and also with the General to Specific (GETS) approach.

Keywords:

• demand for M1
• US
• structural breaks
• income elasticity
• cost of holding money

JEL Classification:

• E41

Notes

¹ Studies by Lucas (1988), Poole (1988), Stock and Watson (1993) and Hoffman et al. (1995) asserts that the demand for M1 in US is stable over the twentieth century. However, Goldfeld (1976) and Judd and Scadding (1982) found that the money demand is unstable during 1970s. Further, McNown and Wallace (1992) and Yossifov (1998) obtained implausible income elasticity of M1 and M2. See Sriram (1999) for a survey.

² The pros and cons on whether monetary aggregates or the interest rate should be used as monetary policy instrument are based on whether the effects of monetary policy are through the real balance effect or through interest rate effect. Objections to the use of the rate of interest as policy instrument may also be raised because changes to the rate of interest would have important distributional effects and these are ignored in debates on the merits of using it as a policy instrument.

³ One presumes that their JML estimates were unsatisfactory. The four alternative methods are Stock and Watson's (1993) Static OLS (SOLS) and Dynamic OLS (DOLS), Phillips's (1991) Band Spectral (PBSR) estimator and Phillips and Hansen's (1990) Fully Modified (PHFM) estimator.

⁴ Unfortunately, Choi and Jung's dating the break dates gives the impression that they are using monthly and not quarterly data. Therefore, our notation of the dates for the subsamples based on what they might have intended to.

⁵ Gregory and Hansen used the demand for money to illustrate their techniques and per se they are not interested in a wider sense in the issues on the demand for money. Rao and Kumar (2007, 2009) have used their tests for testing structural breaks in the demand for money of Fiji and Bangladesh.

⁶ A non linear trend or a linear trend with dummy variables can also be included.

⁷ See Baba et al. (1992, p. 29) for a similar reasoning for the inclusion of the rate of inflation. In our sample the correlation between interest rate and inflation rate is about 0.7 implying that 49% of interest changes are explained by inflation.

⁸ Baba et al. (1992) analysed the US demand for money from this perspective and although they did not conduct formally tests for structural breaks, they showed that their improved specification adequately explains the missing money episode (1974 to 1976), great velocity decline (1982 to 1983) and M1 Explosion (1985 to 1986); see Baba et al. (1992, fn. 1) for the details of these episodes.

⁹ To conserve space we report below only the Maximal Eigenvalue test for Equations 1 and 3.

¹⁰ Perhaps this may be the reason why Choi and Jung did not report JML estimates of the cointegrating equations for their subsamples even though JML procedure has been used to test for cointegration. In our JML estimates for the subsample all the estimated coefficients are insignificant but are correctly signed. In the second subsample the coefficients of income and trend are wrongly signed and all are insignificant.

¹¹ Perhaps the exception is Baba et al. (1992) who estimated with quarterly data (1960Q3 to 1988Q3) demand for money in the US. Baba et al. (1992, p. 26) observe that ‘We infer that the reason for the shifts in alternative models is their omission of appropriate dynamic structure and of important variables’.

¹² Generally it is mistaken that if the adjustment coefficient exceeds unity there is no convergence. However this is valid only if the absolute value of the coefficient exceeds 2 but there would be fluctuations in the adjustment path if the estimate is unity or more than unity but below 2.

¹³ These errors are in 1963, 1966, 1971, 1974 to 1978, 1987, 1989, 1993, 1996, 1998 to 1999, 2002, 2004 and 2008.