The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
FEDERAL RESERVE BANK of ATLANTA Inflation and Monetary Regimes Gerald P. Dwyer and Mark Fisher Working Paper 2009-26 September 2009 WORKING PAPER SERIES FEDERAL RESERVE BANK o f ATLANTA WORKING PAPER SERIES Inflation and Monetary Regimes Gerald P. Dwyer and Mark Fisher Working Paper 2009-26 September 2009 Abstract: Correlations of inflation with the growth rate of money increase when data are averaged over longer time periods. Correlations of inflation with the growth of money also are higher when high-inflation as well as low-inflation countries are included in the analysis. We show that serial correlation in the underlying inflation rate ties these two observations together and explains them. We present evidence that averaging increases the correlation of inflation and money growth more when the underlying inflation rate has higher serial correlation. JEL classification: E31, E5 Key words: money and inflation, inflation, quantity theory The authors have benefited from comments from participants in the conference Money and Monetary Policy, sponsored by the Federal Reserve Bank of Atlanta and Fordham University. James R. Lothian and Paul D. McNelis provided helpful comments on an earlier draft. Budina Naydenova provided helpful research assistance. The views expressed here are the authors’ and not necessarily those of the Federal Reserve Bank of Atlanta or the Federal Reserve System. Any remaining errors are the authors’ responsibility. Please address questions regarding content to Gerald P. Dwyer, Research Department, Federal Reserve Bank of Atlanta, 1000 Peachtree Street, N.E., Atlanta, GA 30309-4470, 404-498-7095, 404-498-8810 (fax), jerry@jerrydwyer.com, and University of Carlos III, Madrid, Departamento de Economía de la Empresa, Calle Madrid, 126,28903 Getafe, Madrid, Spain, or Mark Fisher, Research Department, Federal Reserve Bank of Atlanta, 1000 Peachtree Street, N.E., Atlanta, GA 30309-4470, 404-498-8757, 404498-8810 (fax), mark.fisher@atl.frb.org. Federal Reserve Bank of Atlanta working papers, including revised versions, are available on the Atlanta Fed’s Web site at www.frbatlanta.org. Click “Publications” and then “Working Papers.” Use the WebScriber Service (at www.frbatlanta.org) to receive e-mail notifications about new papers. 1 Introduction Is inflation related to money growth? Many have interpreted recent low correlations of money growth and inflation as evidence that inflation is not related to money growth under all circumstances, perhaps especially in low-inflation environments. There is a large literature showing that money growth and inflation are related. The earliest papers in modern times were associated with the Money and Banking workshop at the University of Chicago (Friedman 1956). While Anderson and Jordan’s (1968) paper using quarterly data for the United States was controversial, it clearly showed that money and inflation were related for 1952 to 1968. More recent papers suggest that such a relationship is not as close or as informative since the decline in inflation in the U.S. in the 1980s. Kishor and Kochin (2007) show that part if not all of the explanation for the change in the importance of money growth in the United States is the change in the importance attached to inflation in monetary policy. When the monetary authority targets inflation using a control variable, the simple relationship between inflation and the control variable will decline because the control variable is changing to offset other influences on inflation. Kishor and Kochin show that the evidence for the United States is quite consistent with this analysis and an increasing emphasis on stabilizing inflation in U.S. monetary policy. Empirical results across countries are not unequivocal either. Lucas (1980), Lothian (1985), Dwyer and Hafer (1988, 1999), McCandless and Weber (1995), Rolnick and Weber (1997) and others find substantial correlations of money growth and inflation across countries for different time periods. Moroney (2002) and De Grauwe and Polan (2005) examine a common criticism of such analyses, namely that the correlations are driven by the high inflation countries and there is little relationship between money growth and inflation for low inflation countries. Moroney (2002) selects countries based on money growth rates and finds a positive relationship between money growth and inflation in low-money-growth countries, but the relationship is stronger and more striking when countries with higher money growth are included in the analysis. De Grauwe and Polan present evidence that the correlations are close to zero or zero for low inflation countries. Frain (2004), responding to the 2001 workingpaper version of De Grauwe and Polan’s paper, removes countries with visible, documented data discontinuites or less than 25 years of data. He finds nonzero correlations of inflation and money growth relative to real income growth for low inflation countries as well as high inflation countries. He also finds regression coefficients for low inflation countries that are not different than 1 one at the five percent significance level. 1 Lothian and McCarthy (2009) proceed in a different way, comparing differences in growth rates across periods with evident differences in inflation. Even though the mean inflation rate in the high-inflation regime is less than ten percent, they find a close connection between the increase in inflation to these low levels and an increase in the the growth of money relative to real income. The length of time over which growth rates are computed has an important influence on the analysis as well. Dwyer and Hafer (1988, 1999) show that the relationship across countries is not particularly obvious over periods as short as a year and is unambiguous over five-year periods. McCandless and Weber (1995) use data over 30 year periods to analyze the relationship and find clear relationships. De Grauwe and Polan (2005) find a relationship when using all countries over a 30-year period as does Frain (2004) for a 25-year period. While interesting and possibly informative, if the only reliable relationship between money growth and inflation is over a quarter of century, that certainly is very long run. 2 Some Evidence on Money Growth and Inflation Before proceeding to our analysis, we document the results discussed above, including the importance of averaging over time and the implications of using only low money-growth countries. First, there is a noticeably closer relationship between inflation and money growth over longer periods than shorter periods. This is at least as strong a characteristic of the data as the other observation: countries with relatively high money growth show this relationship more clearly and make a substantial contribution to the apparent relationship across countries. We also summarize some empirical results about coefficients in regressions of inflation on money growth. Throughout this paper, we measure the nominal quantity of money and its growth rate relative to real income. Adjusting the nominal quantity of money in this manner is useful if the income elasticity of the demand for money is unity or not too far from unity. While this is not a particularly important adjustment when there is substantial variation in inflation relative to real income growth, it is more important when inflation variation is on the order of magnitude of the variation in inflation. [Insert Figure 1 about here] 1 The standard errors are larger for low-inflation countries, but confidence intervals include one and do not include zero. 2 The price level and money relative to real income are strikingly similar for both of the two high inflation countries shown in Figure 1, Brazil from 1912 to 2006 and Chile from 1940 to 2006. The price level is the Gross Domestic Product (GDP) deflator. The measure of money is the nominal quantity of money divided by real GDP. All of the series are set to have average values of 100 for the time periods covered. The vertical axis is a proportional scale, making it possible to read growth rates from the slopes of the lines. The closeness of the behavior of the price level and money in both graphs, including the decreases in inflation toward the end of the periods, is striking. [Insert Figure 2 about here] Figure 2 shows similar graphs for two countries with relatively low inflation, Japan and the United States for the parts of the postwar period with consistent data. The movements of the price level and money relative to real income are similar, but definitely not as close as in the graphs for Brazil and Chile. As earlier papers indicate, there is little obvious short-term relationship. In the United States, money relative to income is low compared to the price level in the 1990s. The fall in the price level in Japan from 1998 to 2006 is associated with lower growth of money relative to income for those years, but there is no corresponding fall in the level of money relative to income. Money relative to income increases 0.8 percent per year from 1998 to 2006 while prices fall 1.3 percent per year. [Insert Figure 3 about here] Figure 3 shows average inflation rates and growth rates of money relative to real income across countries. As does Frain (2004), we use the term “excess money growth” instead of the more cumbersome “growth of money relative to real income”. The upper left panel shows the relationship between the two for all 166 countries for which we have data for twelve or more consecutive years. We include data starting in 1985 or later through the end of the period. Not all of the countries have data for the whole period, for example Albania’s data begin in 1994. The upper right panel shows the relationship for countries with an excess money growth rate less than 50 percent, 159 of the countries. The lower left panel shows the relationship for countries with excess money growth less than 20 percent, and the lower right panel shows the relationship for countries with excess money growth less than 10 percent. We use the growth of money relative to real income instead of the inflation rate to pick countries with low inflation because regressions of inflation on money growth have biased coefficients if countries are picked on the basis of the dependent variable, the inflation rate. 2 2 Suppose that inflation and the growth of money relative to real income are related with a coefficient of one and money growth is exogenous to inflation. If inflation is used to pick countries, then the dependent variable is being used to select the 3 For all of the countries, there is a positive relationship between inflation and excess money growth. We find a positive relationship for low inflation countries which we define as those with excess money growth less than ten percent per year. The correlation monotonically decrease with decreases in the cutoff growth of excess money but it does not go to zero. The correlation is 0.47 even for countries with the average growth of excess money less than ten percent. Figure 3 also shows lines for regressions of inflation on excess money growth. The slopes in these regressions also are used by some as a criteria for evaluating the usefulness of money as a predictor of inflation, with coefficients close to one being considered more supportive (Moroney, 2002; Frain, 2004; DeGrauwe and Polan, 2005). As the data are cut off at lower growth rates of excess money, the regression coefficients decrease, with the regression coefficients in this figure decreasing from 1.01 for all the data, to 0.99 for countries with excess money growth less than 50 percent, to 0.88 for countries with excess money growth less than 20 percent and to 0.41 for countries with excess money growth less than 10 percent. The regression coefficient is substantially less than unity for lower growth rates of excess money. Kisher and Kochin’s (2007) analysis suggests why this is so. The correlation of inflation and excess money growth is zero if all deviations from a constant target inflation rate are unpredictable. The fall in the correlation and the regression coefficient is consistent with their analysis if low inflation countries have less variability of inflation targets and therefore less correlation of inflation and excess money growth. [Insert Figure 4 about here] Figure 4 shows the relationship between inflation and money growth when the data are averaged over successively shorter periods. The upper left panel shows the relationship over all the years for which we have data on each country, which is as much as 21 years and as few as twelve years. The upper right panel shows the relationship with data averaged over the last ten years for which we have data. The lower panels show the relationship with data averaged over five years and one year. It is clear that the relationship becomes weaker over shorter periods. This is consistent with averages presented over five years and less presented by Dwyer and Hafer (1988, 1999). Figure 4 also shows regression coefficients of inflation on excess money growth. These coefficients also decrease as the data are averaged over shorter periods, observations. Countries with inflation greater than 10 percent and growth of money relative to real income less than 10 percent are excluded but countries with inflation less than 10 percent and growth of money relative to real income greater than 10 percent are included. This selection biases the regression coefficient downward from one. 4 from 1.01 for all the data to 0.46 for one year of data. 3 Money Growth and Inflation Why are money growth and inflation more closely related when data are averaged over long time periods and when high inflation countries are included in an analysis of inflation and money growth? In this section, we provide an explanation based on variation of the underlying inflation rate relative to the demand for money. This analysis predicts that higher serial correlation of the underlying inflation rate is associated with a larger increase in the correlation between inflation and money growth as more years are averaged. We also show that the size of the slope coefficient in a regression of inflation on money growth is uninformative about whether the quantity theory holds. The quantity theory is consistent with a slope coefficient of unity in a regression of inflation on money growth and it is consistent with a slope coefficient less than unity. Suppose that the demand for money has unit income elasticity and no other variables systematically affect demand. Then μt − yt = π t + εt (1) where μt is the growth rate of the nominal quantity of money in period t, π t is the inflation rate, yt is the growth rate of real income and εt is an error term in the demand for money. 3.1 Inflation Targeting Suppose that the monetary authority’s actions target the inflation rate, whether this is intentional or not, and the target is π ∗t which varies over time. This relationship can be written π t = π∗t + ηt , (2) where η t is the error term in this equation. For simplicity, we suppress the subscript t. Combining (1) and (2) results in μ − y = π + ε = π ∗ + η + ε. (3) The correlation of the inflation rate (2) and excess money growth (3) is ρ = Corr [π, μ − y] = Cov [π ∗ + η, π ∗ + η + ε] , SD [π ∗ + η] SD [π ∗ + η + ε] 5 (4) which equals ρ= Var [π ∗ + η] + Cov [ε, π ∗ + η] SD [π ∗ + η] (Var [π ∗ + η] + 2Cov [ε, π ∗ + η] + Var [ε])1/2 . (5) At first glance, it is not obvious this is particularly helpful. Suppose, though, that the error term in the demand for money is orthogonal to the target price level and errors in hitting it, i.e. Cov [ε, π ∗ + η] = 0. Then the correlation of the inflation rate and excess money growth simplifies to ρ= SD [π ∗ + η] (Var [π ∗ + η] + Var [ε])1/2 (6) . This equation for the correlation can be interpreted in an informative way. First off, suppose that the variance of the inflation target and errors in hitting it are zero. Equation (6) states the obvious: the correlation of the inflation rate with excess money growth is zero. If there is substantial variance in inflation targets or errors in generating that inflation rate relative to the demand for money, then the correlation will be closer to one. This is related to the analysis by Kishor and Kochin (2007); it also provides a tenative explanation of Figure 3. For countries with similar inflation targets, i.e. little variance of inflation targets or errors in hitting them, the correlation across countries of inflation with excess money growth will be low if not zero. At the other end of the range between zero and one, zero variance of the error term in the growth of money demand implies Corr [π, μ − y] = 1. (7) The demand for money and the monetary authority’s inflation target may well have different characteristics over time. Suppose that the inflation target varies gradually over time and the demand for money varies more over short periods of time. Then the relative variance of π ∗ + η and ε will change as data are averaged over different time periods. Over short periods, the variance of the demand for money will be larger relative to the variance in the supply; over longer periods, the variance of the demand for money decreases relative to the variance in the supply. In the limit, the variance in demand goes to zero and the correlation of the inflation rate with excess money growth goes to one. Orthogonality of the error term in the demand for money and the error in the supply of money and changes in the target inflation rate is sufficient for this characterization of the correlations but is not necessary. The correlation can be written ρ= 1 + Cov [ε, π] /Var [π] (1 + 2Cov [ε, π] /Var [π] + Var [ε] /Var [π])1/2 6 . (8) Even if Cov [ε, π] 6= 0, this correlation approaches one as Var [π] increases relative to Cov [ε, π] and Var [ε] . In short, a higher correlation of inflation and excess money growth is to be expected with a higher variance of the inflation target and errors hitting it if the covariance of errors in the growth of money demand with the inflation rate and the variance in errors in the demand for money do not increase proportionately. The regression coefficient from a regression of inflation on excess money growth will not be unity even though the quantity theory holds in this setup. This regression coefficient is β π|μ−y = Cov [π, μ − y] , Var [μ − y] (9) which can be rewritten as β π|μ−y = Var [π] Cov [π, ε] + . (10) Var [π] + 2Cov [ε, π] + Var [ε] Var [π] + 2Cov [ε, π] + Var [ε] This does not obviously equal one, and it does not equal one in general. Even if Cov [ε, π] = 0, Var [π] , (11) β π|μ−y = Var [π] + Var [ε] which is less than one unless Var [ε] is zero. 3 Stated more positively, β π|μ−y approaches one as Var [ε] /Var [π] goes to zero but, with inflation targeting, the coefficient does not equal one even if the covariance of errors in the growth of money demand and the inflation rate is zero. This result does not hold under all circumstances. 3.2 Control of Money Supply Instead of being determined by the demand for money as it would be under inflation targeting, suppose the supply of money is determined by μ = π∗ + y + ζ, (12) where ζ is the error term and the demand for money is the same as equation (1). The money supply is determined with a target inflation rate as the goal but the growth rate of the nominal quantity of money is changed to effect the 3 The similarity of this formula and the one for regressions with errors in the right-hand-side variables is not an accident. With inflation targeting, shocks to the demand for money affect the growth of the nominal quantity of money but not the inflation rate. This is similar to measurement error in a right-hand-side variable that has no effect on a left-hand-side variable. 7 goal. The growth rate of real income is included in the equation for the supply of money with a coefficient of one to reflect the growth of demand due to real income. The central bank can achieve its target inflation rate by changing the growth rate of the nominal quantity of money with the growth rate of real income. This equation (12) can be rewritten μ − y = π ∗ + ζ. (13) Equating the growth of the demand for the nominal quantity of money (1) and the supply of the nominal quantity of money (13) yields π = π ∗ + ζ − ε. (14) It follows that the correlation of the inflation rate and excess money growth ρm is ρm = 1 − Cov [ε, π ∗ + ζ] /Var [π ∗ + ζ] (1 − 2Cov [π ∗ + ζ, ε] /Var [π ∗ + ζ] + Var [ε] /Var [π ∗ + ζ])1/2 . (15) If Cov [ε, π ∗ + ζ] = 0, then ρm = SD [π∗ + ζ] (Var [π ∗ + ζ] + Var [ε])1/2 = 1 (1 + Var [ε] /Var [π ∗ + ζ])1/2 (16) which approaches one as Var [ε] /Var [π ∗ + ζ] goes to one. This is the same conclusion as above under inflation targeting. The conclusion concerning regression coefficients does change though. The coefficient from regressing the inflation rate on excess money growth is β ∗π|μ−y = Cov [π, μ − y] Var [π ∗ + ζ] − Cov [ε, π ∗ + ζ] = Var [μ − y] Var [π∗ + ζ] Cov [ε, π ∗ + ζ] =1− . Var [π ∗ + ζ] (17) If Cov [ε, π ∗ + ζ] = 0, then β ∗π|μ−y = 1. This is not true if Cov [ε, π ∗ + ζ] 6= 0, although β ∗π|μ−y approaches one as Cov [ε, π ∗ + ζ] /Var [π ∗ + ζ] goes to zero. In sum, if the covariance of the errors in the demand for money and supply of money is zero, the correlation of inflation and excess money growth increases to one as the variance in the demand for money goes to zero relative to the the variance in the supply for money. This conclusion concerning the correlation’s value holds whether the monetary regime is one of inflation targeting or control of the money supply. The coefficient in a regression of inflation on the excess money growth rate depends on how the nominal quantity of money is determined. 8 4 Persistence in the Underlying Inflation Rate In this section, we derive testable predictions concerning the evolution of the underlying inflation rate and the correlation of excess money growth and inflation. We show that serial correlation of the underlying inflation rate is consistent with increases in the correlation of inflation and excess money growth as data are averaged over longer periods, and higher serial correlation is consistent with a greater increase in the correlation. Consistent with the argument above, suppose that errors in the demand for money are serially uncorrelated but suppose that the underlying inflation rate, or inflation target, evolves over time according to π∗ = βπ ∗−1 + ν, (18) h i where ν is serially uncorrelated as well and 0 ≤ β < 1. Assume Cov π ∗−1 , ν = 0. If Cov [ε, π] = 0, the earlier analysis show that the one-period correlation of inflation and excess money growth is ρ= SD [π] (Var [π] + Var [ε])1/2 (19) . Let π2 = (π + π−1 ) /2 and ε2 = (ε + ε−1 ) /2. Given these definitions, the correlation of two-period averages of inflation and excess money growth is ρ2 = SD [π2 ] (Var [π 2 ] + Var [ε2 ])1/2 (20) . Is ρ2 > ρ? Because Var [π 2 ] = 12 Var [π] + 12 Cov [π, π −1 ] and Var [ε2 ] = 12 Var [ε] , " Var [π] + Cov [π, π−1 ] ρ2 = Var [π] + Cov [π, π −1 ] + Var [ε] #1/2 . (21) Given serially uncorrelated errors that are mutually uncorrelated, Cov [π, π −1 ] = βVar [π ∗ ] and therefore " Var [π] + βVar [π ∗ ] ρ2 = Var [π] + βVar [π ∗ ] + Var [ε] #1/2 . (22) The issue is whether ρ2 R ρ. This is equivalent to deciding whether β Var [π ∗ ] Var [ε] R 0. Var [π] Var [π] 9 (23) (24) Since this is positive by assumption, it follows that (25) ρ2 > ρ1 . This algebra makes it clear that the correlation goes up with averaging as ∗ ] Var[ε] β Var[π increases. The parameter β represents the serial correlation in Var[π] Var[π] the underlying inflation rate. Increases in β increase the difference between the one-period and two-period correlations. If β = 0, then ρ2 = ρ1 . If β > 0, then ρ2 > ρ1 . The other terms represent the product of the variation in the inflation target relative to the inflation rate and the variation in the demand for money relative to the inflation rate, which affect the magnitude of the increase in the correlation. 5 The Generalized Local Level Model We can examine the relationship between the correlations and the serial correlation of the underlying inflation rate π ∗ in a state-space model in which π ∗ is unobservable. We use a Bayesian analysis to derive the posterior distribution of the serial correlation parameter β and the relationship between β and the correlation between inflation and excess money growth as the data are averaged over longer periods. The generalized local level model is a simple state-space model involving an observation (measurement) equation and a state (transition) equation. 4 We have T observations on inflation π = (π 1 , . . . , πT ). The state variable is the underlying inflation rate π ∗ = (π ∗1 , . . . , π∗T ) which is not directly observed. The observation and state equations are (for t = 1, . . . , T ) π t = π ∗t + η t π ∗t = δ (1 − β) + β π ∗t−1 + ut , where ⎡ ⎤ ⎢η t ⎥ iid ⎣ ⎦∼ ut N(0, Σ), (26) (27) ⎡ 1 where Σ = h−1 ⎢ ⎣ 0 ⎤ 0⎥ ⎦. ψ (28) We impose the restriction −1 < β ≤ 1. The local level model itself as in Koop (2003) is characterized by β = 1. It is important to deal with the unobserved observation in period 0 in a clean way. Let π ∗0 = δ + w, where w ∼ N(0, h−1 λ) and w is independent of the other 4 This model is a generalization of the local level model given in Koop (2003). The generalization allows the unobserved state variable to be stationary. 10 disturbances. Eliminating π ∗0 from the state equation for π∗1 produces π ∗1 = δ + β w + u1 . (29) Consequently, we see that π ∗1 ∼ N(δ, h−1 (β 2 λ + ψ)). (Note that if β = 0, then λ does not appear in the distribution for π ∗1 .) Given this setup, we now show that π follows a restricted ARMA(1,1) for t ≥ 2. We can eliminate the unobserved state variable and obtain ⎧ ⎨δ + β w + u1 + η1 πt = ⎩ δ (1 − β) + β π t−1 + ut + η t − β ηt−1 t=1 t ≥ 2. (30) Note E0 [w] = E0 [ut ] = E0 [η t ] = 0. Therefore, E0 [πt ] = δ for all t ≥ 1. Define ω t := ut + ηt − β ηt−1 . Let γ ω (τ ) denote the autocovariance function for ω. Then γ ω (0) = ψ h−1 + (1 + β 2 ) h−1 , γ ω (1) = −β h−1 , and γ ω (τ ) = 0 for τ ≥ 2. This autocovariance function is characteristic of an MA(1). As such, we can reexpress ω t as ωt = vt − ξ vt−1 , where |ξ| < 1. 5 We see that π is a restricted ARMA(1,1) for t ≥ 2. If β = 1, then π is an IMA(1,1). The posterior distribution for the unobservables conditional on the observable series π is p(π ∗ , h, φ|π) = p(π ∗ , h|π, φ) p(φ|π), (31) where φ := (δ, β, λ, ψ). (32) The factorization on the right-hand side of (31) will prove convenient. The posterior distribution (31) can be obtained from the joint distribution as follows: p(π ∗ , h, φ|π) ∝ p(π, π ∗ , h, φ) = p(π, π ∗ , h|φ) p(φ) = p(π|π ∗ , h) p(π ∗ |h, φ) p(h|φ) p(φ). (33) According to this parameterization of ω, γ ω (0) = (1 + ξ 2 ) σ 2v and γ ω (1) = −ξ σ 2v . Therefore, q 1 + β 2 + ψ − (1 + β 2 + ψ)2 − 4 β 2 ξ= 2β q 1 + β 2 + ψ + (1 + β 2 + ψ)2 − 4 β 2 2 . σv = 2h 5 Thus the local level model imposes a restriction between ξ and β. Consequently not all ARMA(1,1) processes can be expressed as a generalized local level model. Moreover, β 6= ξ as long as β = 0. Therefore the problem of local non-identification due to cancelation of common factors is absent. 11 The observation and state equations provide p(π|π ∗ , h) and p(π ∗ |h, φ) respectively. To complete the model we must specify the priors p(h|φ) and p(φ). We defer consideration of p(φ) until later. Let the prior for h be independent of φ: p(h|φ) = p(h), with p(h) given by the Gamma distribution: 6 h ∼ G(s−2 , ν). (34) This prior for h delivers analytical expressions for the conditional posterior p(π ∗ , h|π, φ) and for the marginal likelihood p(π|φ). In order to derive the aforementioned analytical expressions for the conditional posterior and marginal likelihood, it is convenient to change the parametrization. As a preliminary, stack the observation equations as follows: π = π∗ + η. (35) Using π ∗ = W θ, we can write (35) as (36) π = Wθ + η where ⎡ θ := ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣ ⎤ π ∗1 ⎥ ⎥ π∗2 − β π ∗1 ⎥ ⎥ ⎥ ⎥ .. ⎥ . ⎥ ⎦ (37) π ∗T − β π ∗T −1 and W is the T × T matrix such that ⎧ ⎨β i−j i≥j otherwise. Wij = ⎩ 0 (38) (Note |W | = 1.) As long as we condition on φ, we can treat β–and hence W –as known. The advantage of the new parametrization appears in the state6 We adopt the parametrization of the Gamma distribution given by Koop (2003, p. 326): If z ∼ G(δ, ν), then the density of z is given by ( ¡ −ν z ¢ ν−2 2 exp c−1 0<z<∞ G z 2δ fG (z|δ, ν) = 0 otherwise, where cG = µ 2δ ν ¶ν 2 Γ ³ν ´ 2 is independent of z and Γ(·) is the Gamma function. Note E[z] = δ and Var[z] = 2 δ 2 /ν. 12 dynamics: ⎧ ⎨δ + β w + u1 θt = ⎩ δ (1 − β) + ut t=1 t > 1. (39) The distributions for π and θ implied by (36) and (39) are ³ π | (θ, h, φ) ∼ N W θ, h−1 IT ³ ´ θ | (h, φ) ∼ N θ, h−1 V , where ⎡ θ= ⎤ δ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢δ (1 − β)⎥ ⎢ ⎥ ⎢ ⎥ .. ⎢ ⎥ . ⎢ ⎥ ⎣ ⎦ ´ (40a) (40b) ⎡ ⎤ 2 > ⎢β λ + ψ 0T −1 ⎥ and V = ⎣ ⎦. 0T −1 ψ IT −1 δ (1 − β) (41) Equation (36) can be interpreted as a normal linear regression model for which (40b) and (34) form a natural conjugate prior: (θ, h)|φ ∼ NG(θ, V , s−2 , ν). (42) Given the conjugacy of the prior (42) with respect to the likelihood (40a), the posterior inherits the form of the prior; namely, (θ, h) | (π, φ) ∼ NG(θ, V , s−2 , ν), (43) where 7 ³ θ = V V −1 θ + W > y ³ V = V −1 + W > W ν =ν +T ³ ´ (44) ´−1 ´> ³ (45) ´ ³ ´> ³ ´ (46) V −1 θ − θ . (47) p(π ∗ |π, h, φ) = p(θ|π, h, φ)|θ=W −1 π∗ = N(W −1 π ∗ | θ, h−1 V ) = N(π ∗ | π ∗ , h−1 V π∗ ), (48) ν s2 = ν s2 + y − W θ y − Wθ + θ − θ The conditional posterior for π ∗ is given by where π∗ = W θ and V π∗ = W V W > . Thus we have established the posterior distribution for π ∗ and h conditional on φ: (π ∗ , h) | (π, φ) ∼ NG(π∗ , V π∗ , s−2 , ν). 7 See Koop (2003, p. 187). 13 (49) We now turn to the posterior for φ. The marginal posterior for φ can be expressed as p(φ|π) ∝ p(π|φ) p(φ). (50) The marginal likelihood of φ is 8 p(π|φ) = ZZ p(π, θ, h|φ) dθ dh = c where ³ Γ(ν/2) ν s2 Ã |V | |V | !1/2 ³ ν s2 ´−ν/2 , (51) ´ν/2 . (52) Γ(ν/2) π T /2 (Note c is free of φ.) The first equality in (51) is identically true while the second equality follows from the specific functional forms given in (34) and (40). c= The parameter space for φ = (δ, β, λ, ψ) is Φ = (−∞, ∞) × (−1, 1] × (0, ∞) × (0, ∞). (53) We adopt the following prior: p(φ) = (2 π)−1 e−( 2 δ 1 2 +λ+ψ) . (54) In addition, values for (s2 , ν) must be specified. We adopt a noninformative prior, setting ν = 0 and setting s2 to an arbitrary value (since s2 enters the posterior only via the product ν s2 ). We can make draws of φ from p(φ|π) via the symmetric random-walk Metropolis MCMC algorithm. The algorithm produces a sequence of draws {φ(r) }R r=1 . Given φ(r) one computes φ(r+1) as follows: Draw φ0 ∼ N(φ(r) , Ω) and u ∼ U(0, 1), and then set (r+1) φ ⎧ ⎨φ0 = ⎩ (r) φ p(φ0 |π) p(φ(r) |π) ≥u otherwise. (55) Equation (55) shows that if the proposal φ0 is “uphill” from the current point φ(r) in the sense that p(φ0 |π) ≥ p(φ(r) |π), then it is always accepted (i.e., added to the output sequence); by contrast, if the proposal is “downhill” from the current point, then it is accepted with a probability that is proportional to the likelihood ratio. (If the proposal is out of bounds, i.e. if φ0 6∈ Φ, then p(φ0 |π) = 0 and the proposal is never accepted.) Note that if the proposal is not accepted, then the current point φ(r) is placed again in the output sequence, which (among other things) induces serial correlation in the sequence of draws. To make the algorithm operational, one chooses a starting value φ(0) ∈ Φ and the 8 See Koop (2003, p. 189) 14 covariance matrix Ω. (The covariance matrix provides a scale for the randomwalk step size.) One must also specify the number of burn-in draws to discard and the amount of thinning to do (if any) after the burn-in period. 9 6 Our Estimates We have a dataset with 166 countries, each of which has continuous annual data for twelve or more years. The number of observations on inflation in each country ranges from twelve to 21. While it would be even better to compare across monetary regimes within a country, the paucity of observations implied by such a strategy leads us to examine the data across countries. 6.1 Estimates of the Serial Correlation Let φi denote the parameters for the i-th country. The parameter of interest is {β i }. We make draws of φi from p(φi |π i ) via the symmetric random-walk Metropolis MCMC algorithm. For each country we make two runs, using the results of the first to calibrate the second. For the first run, we adopt a starting (0) value of φi = (0, .5, 1, 1) and we use Ω = diag(10−4 , 10−2 , 10−2 , 10−2 ) as the covariance matrix for the scale of the step size. We make 104 burn-in draws and then make 105 draws, keeping every one in 102 . Next we compute the mean and covariance of the 103 draws produced by the first run. For the second run, (0) we set φi to the computed mean (from the first run) and we set Ω to .2 times the computed covariance matrix. We make 104 burn-in draws for the second run and then make 105 draws, keeping every one in 102 . This produces a total of 103 draws for each country to approximate the posterior distribution of φi . The draws are not independent and the average first-order autocorrelation across all parameters and all countries is .13. There are only four countries for which the maximum autocorrelation for any of the four parameters in φi is above .5. An approximation for the effective number of independent draws is given by 1−ρ e ≈R , (56) n 1+ρ where R is the number of draws and ρ is the first-order autocorrelation. In e averages about 790 and is not less than 340. our case R = 103 . For β, n The posterior means and the 90% highest posterior density regions for β i are shown in Figure 5. The point estimates for all but 12 countries are positive and the regions do not include zero for 96 of the 166 countries. 9 Koop (2003, Section 5.5) provides a detailed summary of the Metropolis MCMC. 15 [Insert Figure 5 about here] 6.2 Relation to Correlations We are interested in the relation across countries between the posterior means of the autoregressive coefficients and the correlations of the inflation rate and the growth rate of exces money. Let β i denote the posterior mean of β i and ρis denote the sample correlation between the inflation rate and the excess money growth rate for country i over s periods. We summarize the relationship by the correlation between β = (β 1 , . . . , β n ) and ρs = (ρ1s , . . . , ρns ) by the simple correlation rs . Figure 6 shows scatterplots for ρis versus β i for s ∈ {1, 3, 5}. The top row shows plots for all 166 countries, while the bottom row shows plots for only those 119 countries for which we have at least 20 observations. Twenty observations would be considered a small sample for estimating a serial correlation coefficient in most contexts, but requiring 20 observations eliminates over a quarter of these countries. [Insert Figure 6 about here] To estimate the posterior distribution of rs , we apply the Bayesian bootstrap to rs , producing {rs(m) }M m=1 . The Bayesian bootstrap works as follows. Make a (m) draw w from the flat Dirichlet distribution and compute rs(m) using w(m) as probabilities: c(m) rs(m) = q (m) (m) (57) vβ vρs from (m) mβ = n X (m) wi βi (m) vβ = i=1 mρ(m) = s n X n X (m) wi i=1 (m) wi ρis i=1 vρ(m) = s c(m) = n X (m) wi i=1 n X t=1 (m) wi ³ ´ (m) 2 β i − mβ ³ ρis − mρ(m) s ³ (m) β i − mβ ´2 ´³ ´ ρis − mρ(m) . s Figure 7 shows the estimated posterior distributions of the correlations between the serial correlation coefficients and the correlations of inflation and money growth different time spans. All of the rs are positive, indicating that the posterior means of the serial correlation coefficients and the correlation of money growth and inflation clearly are positive. [Insert Figure 7 about here] 16 (m) (m) Next we compare r1 with r3 and r5 . In particular, we compute d3 := r3 − (m) (m) (m) (m) r1 and d5 := r5 − r1 . The results are shown in Figure 8. We find the (m) (m) fraction of d3 that is positive is about .91 while the fraction of d5 that is positive is about .66. If we use only countries that have at least 20 observations, then the fractions increase to about .99 and .93, respectively. These distributions are evidence in favor of the proposition that an increase in the serial correlation coefficient leads to an increase in the correlation of money growth and inflation. [Insert Figure 8 about here] 7 Conclusion The relationship between inflation and excess money growth still is controversial. We find a positive correlation across all countries. The correlation falls as countries with higher excess money growth are excluded, but the correlation is 0.47 across countries with excess money growth of ten percent or less. We show that the lower correlation for low inflation countries is not surprising if low inflation countries have lower variation in unpredictable changes in the supply of money relative to unpredictable changes in the demand for money. We also show that the size of a regression coefficient of inflation on excess money growth is uninformative about the quantity theory. If errors in the supply of money are uncorrelated with errors in the demand for money, then a regression of inflation on the growth rate of money will have a slope coeffient of unity. On the other hand, if this correlation is not zero, as it is with explicit or implicit inflation targeting, the nominal quanity of money is endogenous and a regression of inflation on money growth will not deliver a coefficient of unity. This is perfectly consistent with the quantity theory holding. While regression coefficients equal to unity may seem like a plausible way to evaluate the quantity theory, the quantity theory is consistent with coefficients less than one as well as equal to one. Higher correlations between money growth and inflation when data are averaged over time is consistent with this same analysis. We show that positive serial correlation of the underlying inflation rate is consistent with higher correlations of excess money growth with inflation as the growth rates are computed over longer time periods. We also show that monetary regimes with more sustained deviations of inflation from its mean will show greater increases in the correlation of excess money growth and inflation as the growth rates are computed over longer time periods. 17 We then test these implications. We find substantial variation in serial correlation of underlying inflation rates and we find this variation is positively related to the increase in correlations as data are averaged over longer periods. Our results indicate that sustained excess money growth is positively correlated with inflation. The greater apparence of that relationship when data are averaged over time and when countries with sustained deviations of inflation from its mean inflation are quite consistent with the quantity theory holding. 18 8 Data Appendix We analyze annual data for the United States and for 182 countries. The data across countries include available data for 1985 and subsequent years. These data are from the World Development Indicators website, the March 2008 CD for International Financial Statistics, and from Haver. Haver is the source of the data on Taiwan. The nominal and real Gross Domestic Product (GDP) data primarily are from World Development Indicators. These data are supplemented by data from International Financial Statistics when these IFS data are more complete or consistent. 10 We use the data on money plus quasimoney from IFS except when the WDI data cover a longer period or have more significant digits. 11 The price index is the Gross Domestic Product deflator and nominal income is GDP. Table A1 in the Appendix lists all the countries and the periods over which we have GDP and money data. Data for some individual countries are from Haver or country-specific sources. Data for Taiwan are from Haver because these data are not available in either WDI or IFS. Inspection of some series suggested discontinuities in the underlying data from WDI and IFS. As it turned out, all of the issues concerned the nominal quantity of money. When collecting data from an individual central bank’s website, we collected the monetary series emphasized by the central bank. The nominal quanity of money for Belgium is M3 from the National Bank of Belgium’s website. The quantity of money for Canada is M2 from Haver. The quantity of money for Japan is M2 including certificates of deposit from the Bank of Japan’s website. Earlier and later series are spliced by the average monthly ratio of 0.995519 in the overlapping period April 1998 to March 1999. The quantity of money for New Zealand is M3 from the Bank of New Zealand. The growth rate of the nominal quantity of money for Macedonia is M2 from the Central Bank. The nominal quantity of money for the United Kingdom is M4 from the Bank of England. All data for the United States are from the Federal Reserve Bank of St. Louis’s website and the nominal quantity of money is M2. The long-term data for Brazil and Chile are from Rolnick and Weber (1997) 10 We use some or all IFS data for Anguilla, Aruba, Barbados, Cambodia, Cape Verde, Fiji, Kuwait, Libya, the Maldives, Montserrat, Qatar and San Marino. 11 The WDI data available to us often contain more significant digits when there are large changes in the quantity of money. We use WDI data for Argentina, Bolivia, Brazil, Chile, the Democratic Republic of the Congo, Cyprus, Ethiopia, Ghana, Guinea, Lao People’s Democratic Republic, Lebanon, Malta, Nicaragua, Peru, Romania, Samoa, San Marino, Slovenia, Sudan, Turkey, Turkmenistan, Uganda, Uruguay and Zimbabwe. 19 updated by World Development Indicators. 12 The monetary variable from Rolnick and Weber’s data is their M2. The data are updated by spliced data from the World Development Indicators for 1986 through 2006. 12 Rolnick and Weber (1995) provide the data sources in their Data Appendix. We thank Warren Weber for providing the data. 20 Anderson, L. C., Jordan, J. L. 1968. Monetary and Fiscal Actions: A Test of Their Relative Importance in Economic Stabilization. Federal Reserve Bank of St. Louis Review 50 (11), 11-24. De Grauwe, P., Polan, M. 2005. Is Inflation Always and Everywhere a Monetary Phenomenon? Scandinavian Journal of Economics 107 (2), 239-59. Dwyer, Jr., G. P., Hafer, R. W. 1999. Are Inflation and Money Growth Still Related? Federal Reserve Bank of Atlanta Economic Review 84 (2), 32-43. Dwyer, Jr., G. P., Hafer, R. W. 1988. Is Money Irrelevant? Federal Reserve Bank of St. Louis Review 70 (3), 1-17. Estrella, A., Mishkin, F. S. 1997. Is There a Role for Monetary Aggregates in the Conduct of Monetary Policy? Journal of Monetary Economics 40 (2), 279-304. Frain, J. Inflation and Money Growth: Evidence from a Multi-Country Dataset. Economic and Social Review 35 (3), 251-66. Friedman, M. 1956. Studies in the Quantity Theory of Money. Chicago: University of Chicago Press. Kishor, N. K., Kochin, L. A. 2007. The Success of the Fed and the Death of Monetarism. Economic Inquiry 45(1), 56-70. Koop, G. 2003. Bayesian Econometrics. Chichester, England: John Wiley & Sons Ltd. Lothian, J. R. 1985. Equilibrium Relationships between Money and Other Economic Variables. American Economic Review 75 (4), 828-35. Lothian, J. R. 1976. The Demand for High-Powered Money. American Economic Review 66 (1), 56-68. Lothian, J. R., McCarthy, C. 2009. The Behavior of Money and Other Economic Variables: Two Natural Experiments. Journal of International Money and Finance, this issue. Lucas, Jr., R. E. 1980. Two Illustrations of the Quantity Theory of Money. American Economic Review 70 (5), 1005-14. McCandless, Jr., G. T., Weber, W. E. 1995. Some Monetary Facts. Federal Reserve Bank of Minneapolis Quarterly Review 19 (3), 2-11. Moroney, J. R. 2002. Money Growth, Output Growth and Inflation: Estimation of a Modern Quantity Theory. Southern Economic Journal 69 (4), 21 398-413. Rolnick, A. J., Weber, W. E. 1995. Money, Inflation and Output Under Alternative Monetary Standards. Staff Report 175, Federal Reserve Bank of Minneapolis. Rolnick, A. J., Weber, W. E. 1997. Money, Inflation and Output under Fiat and Commodity Standards. Journal of Political Economy 105 (6), 1308-21. 22 Figure 1 Money and Prices in Brazil and Chile Brazil 1912 to 2006 1000 100 Chile 1940 to 2006 Price level Money relative to real income 100 Percentage of period average 10 10 1 0.1 1 0.01 0.001 0.1 0.0001 1e-005 0.01 1e-006 1e-007 0.001 1e-008 1e-009 0.0001 1e-010 1e-011 1e-005 1e-012 1900 1920 1940 1960 Year 1980 2000 1940 1970 Year 2000 Figure 2 Money and Prices in the U.S. and Japan Japan 1967 to 2007 United States 1959 to 2007 Percentage of period average 2 Price level Money relative to real income 2 102 9 8 102 9 7 8 6 7 5 6 4 5 3 4 2 3 101 2 1970 1990 Year 2010 1971 1982 1993 Year 2004 Figure 3 Inflation and Excess Money Growth for Lower Growth Rates of Excess Money Note: The slope indicated in the figure is the slope of the regression line. The solid line in the figure is the regression line. The dotted line is a regression from the origin with a slope of one. 100 Inflation Rate 100 All countries correlation 0.979 slope 1.006 80 80 60 60 40 40 20 20 0 0 0 Inflation Rate 100 Excess money growth less than 50 percent correlation 0.951 slope 0.992 20 40 60 80 100 0 40 60 80 100 Excess money growth less than 10 percent correlation 0.465 slope 0.411 Excess money growth less than 20 percent correlation 0.853 slope 0.883 80 20 8 60 4 40 20 0 0 0 20 40 60 80 100 Growth Rate of of Excess Money 0 2 4 6 8 10 Growth Rate of of Excess Money Figure 4 Inflation and Excess Money Growth over Shorter Time Periods Note: The slope indicated in the figure is the slope of the regression line. The solid line in the figure is the regression line. The dotted line is a regression from the origin with a slope of one. 80 Inflation Rate 100 All years correlation 0.979 slope 1.006 80 Ten years correlation 0.972 slope 0.844 60 60 40 40 20 20 0 0 0 40 60 80 0 100 Five years correlation 0.916 slope 0.773 Inflation Rate 120 20 200 20 40 60 80 One year correlation 0.807 slope 0.460 150 80 100 40 50 0 0 0 40 80 120 Growth Rate of of Excess Money 0 50 100 150 200 Growth Rate of of Excess Money Figure 5. Posterior means and 90% highest posterior density regions for βi for 166 countries (sorted by mean). 1.0 0.5 0.0 -0.5 -1.0 0 50 100 150 Figure 6. Scatterplots of ρis versus β i for s ∈ {1, 3, 5}. The sample correlations rs are shown. All 166 Countries r1 = 0.38 r3 = 0.44 r5 = 0.40 1.0 1.0 1.0 0.5 0.5 0.5 0.0 0.0 0.0 -0.5 -0.5 -0.5 -1.0 -1.0 -0.5 0.0 0.5 1.0 -1.0 -1.0 -0.5 0.0 0.5 1.0 -1.0 -1.0 -0.5 0.0 0.5 1.0 119 countries with at least 20 observations r1 = 0.37 r3 = 0.49 r5 = 0.47 1.0 1.0 1.0 0.5 0.5 0.5 0.0 0.0 0.0 -0.5 -0.5 -0.5 -1.0 -1.0 -0.5 0.0 0.5 1.0 -1.0 -1.0 -0.5 0.0 0.5 1.0 -1.0 -1.0 -0.5 0.0 0.5 1.0 Figure 7. Bayesian bootstrap distributions of rs . n = 166 and s = 1 n = 166 and s = 3 n = 166 and s = 5 6 6 6 5 5 5 4 4 4 3 3 3 2 2 2 1 1 1 0 0.0 0 0.0 0.2 0.4 0.6 0.8 1.0 n = 119 and s = 1 0.2 0.4 0.6 0.8 1.0 0 0.0 n = 119 and s = 3 6 6 5 5 5 4 4 4 3 3 3 2 2 2 1 1 1 0 0.0 0 0.0 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 0.4 0.6 0.8 1.0 n = 119 and s = 5 6 0.2 0.2 1.0 0 0.0 0.2 0.4 0.6 0.8 1.0 Figure 8. Differences d3 (thick) and d5 . All 166 Countries 119 Countries with T ³ 20 10 10 8 8 6 6 4 4 2 2 0 0 -0.2 -0.1 0.0 0.1 0.2 0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 Table 1 Countries and Years Country First Year Last Year Albania 1994 2006 Algeria 1985 Anguilla Country First Year Last Year Brazil 1985 2006 2006 Bulgaria 1991 2006 1990 2005 Burkina Faso 1985 2006 Antigua and Barbuda 1985 2006 Burundi 1985 2006 Argentina 1985 2006 Cambodia 1993 2006 Armenia 1992 2006 Cameroon 1985 2006 Aruba 1991 2004 Canada 1985 2006 Australia 1985 2006 Cape Verde 1985 2006 Austria 1985 1997 Central African Republic 1985 2006 Azerbaijan 1992 2006 Chad 1985 2006 Bahamas, The 1985 2002 Chile 1985 2006 Bahrain 1985 2005 China 1985 2006 Bangladesh 1985 2006 Colombia 1990 2006 Barbados 1985 2004 Comoros 1985 2006 Belarus 1994 2006 Costa Rica 1985 2006 Belgium 1985 1998 Cote d'Ivoire 1985 2006 Belize 1985 2006 Croatia 1993 2006 Benin 1985 2006 Cyprus 1985 2006 Bhutan 1985 2006 Czech Republic 1993 2006 Bolivia 1985 2006 Denmark 1985 2006 Botswana 1985 2006 Djibouti 1990 2006 Dominica 1985 2006 Haiti 1985 2006 Dominican Republic 1985 2006 Honduras 1985 2006 Ecuador 1985 2006 Hong Kong 1991 2006 Egypt 1985 2006 Hungary 1985 2006 El Salvador 1985 2006 Iceland 1985 2006 Equatorial Guinea 1985 2006 India 1985 2006 Estonia 1991 2006 Indonesia 1985 2006 Country First Year Last Year Ethiopia 1985 2006 Euro Area 1995 Fiji Country First Year Last Year Iran 1986 2006 2007 Israel 1985 2006 1985 2006 Italy 1985 1997 Finland 1985 1997 Jamaica 1985 2006 France 1985 1997 Japan 1985 2006 Gabon 1985 2006 Jordan 1985 2006 Gambia, The 1985 2006 Kazakhstan 1993 2006 Germany 1985 1997 Kenya 1985 2006 Ghana 1985 2006 Kuwait 1992 2006 Grenada 1985 2006 Laos 1987 2006 Guatemala 1985 2006 Latvia 1993 2006 Guinea 1989 2005 Lebanon 1988 2006 Guinea-Bissau 1986 2006 Lesotho 1985 2006 Guyana 1985 2006 Liberia 1991 2006 Libya 1985 2006 Netherlands 1985 1997 Lithuania 1993 2006 New Zealand 1985 2006 Macao 1985 2006 Nicaragua 1985 2006 Macedonia 1992 2006 Niger 1985 2006 Madagascar 1985 2006 Nigeria 1985 2005 Malawi 1985 2006 Norway 1985 2003 Malaysia 1985 2006 Oman 1985 2005 Maldives 1985 2006 Pakistan 1985 2006 Mali 1985 2006 Panama 1985 2006 Malta 1985 2006 Papua New Guinea 1985 2006 Mauritania 1985 2003 Paraguay 1985 2006 Mauritius 1985 2006 Peru 1985 2006 Mexico 1985 2006 Philippines 1985 2006 Moldova 1991 2006 Poland 1990 2006 Mongolia 1991 2006 Portugal 1985 1997 Country First Year Last Year Montserrat 1985 2005 Morocco 1985 Mozambique Country First Year Last Year Qatar 1985 2006 2006 Republic of the Congo 1985 2006 1988 2006 Romania 1985 2006 Myanmar 1985 2005 Russia 1993 2006 Namibia 1990 2006 Rwanda 1985 2005 Nepal 1985 2006 Saint Kitts and Nevis 1985 2006 Saint Lucia 1985 2006 Taiwan 1985 2007 Saint Vincent and the Grenadines 1985 2005 Tanzania 1988 2006 Samoa 1985 2006 Thailand 1985 2006 Saudi Arabia 1985 2006 Togo 1985 2006 Senegal 1985 2006 Tonga 1985 2006 Seychelles 1985 2006 Trinidad and Tobago 1985 2006 Sierra Leone 1985 2006 Tunisia 1985 2006 Singapore 1985 2006 Turkey 1985 2006 Slovakia 1993 2006 Uganda 1985 2006 Slovenia 1991 2006 Ukraine 1992 2006 Solomon Islands 1985 2006 United Arab Emirates 1985 2005 South Africa 1985 2006 United Kingdom 1985 2006 South Korea 1985 2006 United States 1985 2006 Spain 1985 1997 Uruguay 1985 2006 Sri Lanka 1985 2005 Vanuatu 1985 2006 Sudan 1985 2006 Venezuela 1985 2006 Suriname 1985 2006 Yemen 1990 2006 Swaziland 1985 2006 Zambia 1993 2006 Sweden 1985 2006 Zimbabwe 1985 2005 Switzerland 1985 2006 Syria 1985 2006