Examining the time trajectory of GDP growth: a multi country study.

Abstract

In this study, we first present an extensive review of the literature on the time dynamics of GDP, leading up to modern times. We then test for the simultaneous presence (and/or absence) of non-linearity and non-stationarity in the quarterly real GDP growth rate series for the Canada, Japan, United States and the United Kingdom, encompassing the four decades from the 1960's to the present (2004.) Caner-Hansen (CH, 2001) provides a new test which for the first time can simultaneously examine for both non-stationarity and non-linearity (threshold effects).

Existence of a threshold effect would imply that policy intervention occurs only when changes in the real GDP either exceed or are less than a socially and politically acceptable threshold level. This is supported by our results, which indicate the existence of a threshold effect for three of the four countries.

INTRODUCTION

Considerable effort has recently been directed towards examining the cause and effect of the observed volatility reduction in the US GDP over time. Indeed the magnitude of fluctuations of the US GDP has decreased (measured in terms of standard deviation (SD) of quarterly growth rates of real GDP) dramatically since World War II. In the period 1984 onwards (as compared to the post war period 1950-83) the SD has decreased by one half, a significant drop indeed.

GRAPH HERE

This begs an answer to the question as to why this volatility reduction came about. It is also important from the policy maker's point of view. Policy making is crucially dependent on a proper analysis of macroeconomic forecasts, since it is these predictions and their projected path, that help in reducing future uncertainty and volatility in the variable under consideration. Studying the movements of aggregate economic variables in a scientific manner will help in better economic planning over the future periods. But success in modeling and forecasting GDP depends on a proper analysis of the dynamics of this variable itself, i.e., a credible characterization of its time path and the nonstationarities and nonlinearities inherent in its behavior.

The empirical literature in this particular case has not proceeded in a logical step by step manner, which would be to first test for the presence of nonlinearities in the level / growth rate of the GDP series, and then use an appropriate nonlinear model to replicate its time path over business cycles. In most cases than not, nonlinearity in GDP has been assumed and not tested. Here we would like to take a step back, and test for nonlinearities in the rate of growth of the GDP of four major economies (Canada, Japan, USA and UK), using a recently available (and powerful) test, which would account for the first part of the contribution of this exercise.

The second contribution has to do with the standard assumption made in the literature regarding stationarity of the variable under consideration, and then only testing for nonlinearity, in isolation. This fundamental flaw in the econometric / empirical methodology encompasses all the parametric as well as nonparametric studies. They are all limited by their assumption that the series under consideration is stationary (more specifically trend stationary, and hence using the common solution of simply detrending the data), and therefore the test is for nonlinearity only. The problem here is that even if this assumption is correct, this automatic "detrending" before setting up the actual threshold autoregressive model would bias the results. In our context, it becomes all the more relevant since we have overwhelming empirical evidence of nonstationarity (unit roots) in the level of real output, see for example Nelson and Plosser (1982), Wasserfallen (1986), Watson (1986), Cheung and Chinn (henceforth CC, 1996) and Rapach (2002). Whether level stationarity or otherwise (unit roots) is true has an important bearing for the time path of GDP or the trajectory of business cycles. It would also affect our forecasting capability and concomitantly the credibility of macroeconomic stabilization policies.

The empirical evidence overwhelmingly supports non-stationarity of real GDP using both unit root methods (see CC, 1996) or panel unit root tests (see Rapach, 2002). Thus the need is for an empirical test of "linearity" versus "nonlinearity", but without any "a priori" assumptions about the (stationarity/nonstationarity) data series under consideration. This is where the recently available Caner-Hansen (henceforth CH, 2001) study comes into play. It is the first rigorous theoretical and empirical treatment of the simultaneous existence of both nonstationarity and nonlinearity. Now we do not have to assume one to test for the other in isolation, because it allows for nonlinear (asymmetric adjustments) dynamics for unit root testing procedure. This procedure is much more powerful than conventional unit root tests, especially if the adjustment procedure is not symmetric, see Enders and Granger (henceforth EG, 1998).

The third contribution is the "comparative study" being undertaken here of the time dynamics (non-stationarity and non-linearity) of the GDP growth rate of four major world economies namely Canada, Japan, UK and USA. It is of additional interest to examine if these characteristics are uniform (or different) across economies.

This paper proceeds as follows. In section 2 we conduct an extensive and up to date literature review of the relevant articles from both economics and finance, followed by a description of the data used, in section 3. In section 4 we briefly describe the Caner-Hansen methodology. This is followed by an analysis and explanation of the results in section 5, with our concluding remarks in section 6.

LITERATURE REVIEW

An examination of the literature provides ample evidence of the fact that the magnitude of the US GDP fluctuations has decreased dramatically over time. Starting from the early eighties onwards, its volatility (measured in terms of standard deviation of quarterly growth rates of real GDP) has reduced by one half, a significant drop indeed. This begs an answer to the question as to why this reduction came about. An examination of the characteristics / stylized facts of the economy and the nature of the stabilization has been the focus of the study by Ahmed et al., (2002), Blanchard and Simon (2001), Chauvet and Potter (2001), Kim et al., (2004), Kim and Nelson (1999 a, b), McConnell and Perez-Quiros (2000), Niemira and Klien (1994), Stock and Watson (2002) and Warnock and Warnock (2000), and among others. The reasons put forward run the entire gamut of possible rationalizations from improved policy making (fiscal and monetary policy) to structural factors (less friction in the inventory management process and hence better synchronization between demand and supply) to random factors (butterfly effects) to sheer good luck (reduction in the internal and external real shocks.) One school of thought contends it is because aggregate real GDP underwent a reduction in volatility (more due to its cyclical components than trend components), while others find the volatility reduction is pervasive over various sectors of the economy, indicating that this stability was brought about more by improvements in inventory management than by better monetary policy. So the question is, "is it one (or more) factors or a combination of factors?" The answer to this is extremely important, also from the policy maker's point of view.

Examining these fluctuations to see if they can be rationalized by social / economic/ demographic and/or technological factors or if they are purely random is also integrally linked to making credible macroeconomic forecasts. Policy making is crucially dependent on a proper analysis of these forecasts, since it is these predictions (and their path) that help in reducing future uncertainty and volatility in the variable

under consideration. Studying the movements of aggregate economic variables in a scientific manner will help in better economic planning over the future periods. But success in modeling and forecasting GDP depends in an integral manner on a proper analysis of the dynamics of this variable itself, i.e., a credible characterization of its time path and the nonstationarities and nonlinearities inherent in its behavior.

Historically GDP growth (i.e., output changes over time) was studied in a linear set up, following the seminal route prescribed by Box and Jenkins (1976). Works in this line are Braun and Zarnowitz (1989) and Diebold and Rudebusch (1991) among others. But the basic assumption of linearity for business cycles implies symmetric troughs and peaks, which could never be justified by their realistic shape, which were always asymmetric. The reason for the asymmetric shape of expansions and contractions being that the growth rate is faster after a trough but slower after a peak. Economic growth is generally characterized by nonlinear trajectories, fluctuating randomly between two successive periods of expansion and contraction. The reduction in GDP growth volatility is not because of better foresight and less structural friction in the economy, but is more so based on the expectations (and acceptance) of the fact that no regime (expansion or contraction) is permanent. Thus there are built in forces (based on human expectation formation and behavior modification) which turn the tide in the reverse direction. To borrow a terminology from International Finance, "bandwagon effects" in our expectation formation process resulted in lowering the volatility.

Interestingly, the basic reason for assuming that the growth path of the GDP would be nonlinear was outlined way back in 1964 by Milton Friedman, in an NBER study on business cycles. He showed an asymmetry in the correlations between successive business cycle phases, where the amplitude of a contraction and the next expansion were correlated, but the expansion phase was not correlated to the amplitude of the next contraction. He called it the "plucking" model of business cycles. The economics literature studying the path of GDP growth has confirmed this asymmetry time and time again. The asymmetric behavior of business cycles is also evident from the way the production possibility frontier (PPF) is set up. Here recessions are situations below the PPF and thus fluctuations in output are asymmetric in the negative direction. The asymmetric behavior was highlighted by Neftci (1984) who statistically documented the pattern of business cycles. Recessions were consistently found to be steeper than recoveries, as in DeLong and Summers (1986), Falk (1986) and Sichel (1993). It is also evident in the Markov switching model setups used by Hamilton (1989) and Lam (1990 and 2004). Sichel (1993, 1994) found the recessions to be deeper than expansions are tall and reported real GDP to have a "peak reverting" behavior. Sichel conceptualizes them as different levels of steepness and deepness in business cycles during contraction and expansion phases. Then we have Diebold and Rudebusch (1990) and also Durland and McCurdy (1994) who examine business cycle dependence models (in a univariate form) while Kim and Nelson (1998) does so (in a multivariate context.) Another evidence of the same is McQueen and Thorley (1993) who find US business cycles characterized by sharp troughs and round peaks.

In Galvao (2002) the author evaluates a large set of nonlinear univariate time series models of US GDP and finds that nonlinearities are evident due to business cycle asymmetries, which according to Balke and Wynne (1995) are because business cycles tend to be concave during expansions and linear during contractions. Here Galvao uses a linear approximation of the cycle model to trace the actual trajectory and finds that linearization does a poor job, and hence nonlinear approximations are the way to go. Moreover, this study measures the business cycle asymmetries as the "excess" as defined by Harding and Pagan (2002). Using this matrix, it evaluates 15 nonlinear time series models, and finds this "excess" over the lower and upper quartiles of the distribution to be different.

Alternative variants of the threshold autoregressive model have been used. Among them is the Beaudry and Koop (1993), who use an endogenous switching technique to measure the "depth of the recession" which allows for nonlinearity in the model. They use the impulse response technique on real GDP and find that negative innovations are less persistent than positive ones. Extensions on these lines have been done by Bradley and Jansen (1997) and Pesaran and Potter (1997). The latter use three endogenous regimes, namely corridor, floor and ceiling, and try to reproduce the asymmetric shape of the business cycle. Pesaran and Potter also report a nonlinear GDP path with linear contractions and concave expansions. After a thorough comparison, they provide concrete evidence of the importance of nonlinearities in business cycles.")

Then we have Breunig and Stegman (2003) who extend the Bodman and Crosby (1999) Markov Switching (henceforth MS) nonlinear model to study the dynamics of Singapore GDP. This nonlinear setup outperforms the linear AR models. They find that GDP growth is best studied under asymmetric "high growth" and "low growth" phases. Thus the regime switching model is so applicable. They also find the MS model is superior to linear alternatives in "one step ahead" predictions of the future path of GDP. They also use "graphical methods" and demonstrate the added value of using a nonlinear model, like a MS framework. Hence an asymmetric nonlinear setup is an appropriate one.

In recent work by DeJong et al., (2004) the authors build a reduced form model of GDP growth, to study its inherent nonlinearity and volatility. They use a regime switching model between accelerating and decelerating GDP growth rates (also see Kim et al., (2004) who try to capture the pace of output growth between recessions and recoveries, called the bounce back effects.) It follows the Filardo and Gordon (1998) outline where GDP growth accelerates or decelerates stochastically, but there are observable indicators or "tension index" which helps in predicting the probability and time of a regime shift. The authors justify the nonlinearities in the GDP growth path based on its "probable trajectory" and the accompanying rationale for it. At the onset of an expansion, the GDP has a regime drift component that grows/accelerates along an increasing trajectory. This leads to an increase in the "tension index" which enhances the probability of an oncoming regime shift. Once this happens, the process reverses itself, and the drift component begins to decline along a new path. This movement of the drift component between expansion and contraction phases generates a cyclical behavior in the GDP growth rate. Other forms of heterogeneity observed in the dynamics of the GDP time path is that they are all different. No two business cycles are alike, differing either in the magnitude of the cycles (moderate long term growth versus rapid short bursts of growth.)

Given this preponderance of evidence favoring nonlinearities in the basic structure of GDP growth, the rational way to proceed would be to first test for nonlinearity. In most cases, nonlinearity in GDP growth has been assumed and not tested. Here we would like to take a step back, and test for nonlinearity in the US GDP, using a recently available and powerful test, which would account for one part of the contribution of this exercise.

The other has to do with the standard assumption made in the literature regarding stationarity of the variable under consideration, and only testing for nonlinearity, in isolation. This fundamental flaw in the econometric / empirical methodology encompasses all the parametric as well as nonparametric studies. They are all limited by their assumption that the series under consideration is stationary (more specifically trend stationary, and hence the common solution is simple detrending of the data), and therefore the test is for nonlinearity only. The problem here is that even if this assumption (of trend stationarity) is correct, this automatic detrending before setting up the actual threshold autoregressive model would bias the results. (2) Then comes the question "what if the series was a unit root process?" In that case the convergence would have to be justified only in terms of a non-linear adjustment, which actually could (and should) be partly explained by the inherent nonstationarity in the data. Along with that, the true but "undiagnosed" nonstationarity (actual unit roots in the series) would lead to incorrect nonlinear test result inferences, since they all have non-standard asymptotic distributions.

Again, whether level stationarity or otherwise (unit roots) is true has an important bearing for the time path of GDP or the trajectory of business cycles. For example, if real output is trend stationary, then shocks will have only a transitory impact on real output, with convergence happening soon enough. But if it is not so, then real factors such as technology shocks will have permanent ramifications (not fluctuations around a deterministic trend.) Also, it would affect our forecasting capability and credibility of macroeconomic stabilization policies. The empirical evidence overwhelmingly supports non-stationarity of real GDP using both unit root methods (CC, 1996) or panel unit root tests (R, 2002).

Jones (2003) also stressed this flaw but in the context of short term interest rate dynamics. The problem here is that this "a priori" assumption of stationarity imposes restrictions on the possible shape of the drift function, and thus its determined shape itself is suspect. Not only that, but today "nonlinearity in drift" is gaining more and more advocates, mainly because it fits the real life data better. To circumvent these problems, Jones uses a Bayesian approach to study the character of the drift/diffusion process which has exact finite sample properties, even if the process is nonstationary. He finds evidence favoring a nonlinear drift in high frequency data, but concludes that the extent of nonlinearity is dependent on the beliefs / assumptions made about the shape of the drift function. Ahn and Gao (1999) show that a nonlinear drift function improves on fixed income pricing models.

Thus the need for an empirical test of "linearity" versus "nonlinearity", but without any a priori assumptions about the (stationarity/nonstationarity) data series under consideration. This is where the recently available Caner-Hansen (2001) study comes into play. It is the first rigorous theoretical and empirical treatment of the simultaneous existence of both nonstationarity and nonlineraity. Now we do not have to assume one to test for the other in isolation, because it allows for nonlinear (asymmetric adjustments) dynamics for the unit root testing procedure. This procedure is much more powerful than conventional unit root tests, especially if the adjustment procedure is not symmetric, see EG (1998).

DATA

The data is from the OECD (Main Economic Indicators) data set, available from the Estima Corporation. We have obtained quarterly GDP data for Canada, Japan, United Kingdom, and the United States. We then calculated the annualized growth rates of GDP. The time period is Canada [1981 Q1-2004 Q1], Japan [1980 Q1-2004 Q1], United Kingdom [1960 Q1-2004 Q1], and United States [1960 Q1-2004 Q1].

CANER-HANSEN MODEL

In all the studies mentioned here and also in the literature examining regime shifts, the maintained assumption is that the data under consideration is ergodic and stationary. The tests then conducted are for data series nonlinearity and its type. CH (2001) is the first rigorous treatment of the simultaneous existence of both nonstationarity and nonlineraity. There are Wald and "t" test for unit roots, and a sequential Wald test for threshold effects. This null has two components, one reflecting the unit roots but free from nuisance parameters and the other similar to the stationary case but nuisance parameter dependent. Thus the distributions are nonstandardized, and have to be derived in every case. The unit root Wald tests have an asymptotic null distribution that depends on whether or not there is a threshold effect. These tests are superior (more powerful) than conventional ADF tests, when the true process is indeed nonlinear. (3) A standard TAR model is:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (1)

where [e.sub.t] is an i.i.d, error process and [lambda] is an unknown threshold, within the interval [lambda] [member of] [LAMBDA] = [[[lambda].sub.1], [[lambda].sub.2]] where each segment has a significant presence to be dubbed a regime. The i.i.d errors ensure that the first difference of the series [DELTA][y.sub.t] is stationary and ergodic, so that [y.sub.t] is itself integrated of order one. The regimes (i.e, the TAR models) are estimated by least squares.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (2)

The threshold [lambda] is estimated by minimizing [[sigma].sup.2]([lambda]):

[??] = arg mi[begin strike through]n[end strike through] [??]([lambda]), [lambda] [member of] [LAMBDA] (3)

The first difference model:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (4)

is estimated using the standard Wald and "t" statistic. Here the statistics are standard, but the sampling distribution is non standard. The test in eq.(1) is for the presence of threshold effects, under the joint hypothesis [H.sub.0]: [[theta].sub.1] = [[theta].sub.2], implying no regimes. Under the null of no threshold effects, [lambda] is not identified, hence the testing procedure is nonstandard. The null hypothesis of [H.sub.0] : [[theta].sub.1] = [[theta].sub.2] = [theta], simplifies the model to

[DELTA][y.sub.t] = [rho][y.sub.t-1] + [mu] + [alpha]' x [??][y.sub.t-1] + [[epsilon].sub.t] (5)

where [??] [y.sub.t-1] = ([DELTA][y.sub.t-1] ... [DELTA][y.sub.t-k])' are two bootstrap methods, one for stationarity and other for the nonstationary case. In testing for unit roots and nonstationarity CH discuss three possibilities. In eq. 1, [[rho].sub.1] and [[rho].sub.2] are the determinants of stationarity of [y.sub.t]. Under the null hypothesis,

(1) [H.sub.0] [[rho].sub.1] = [[rho].sub.2] = 0, [DELTA][y.sub.t] is stationary, indicating [y.sub.t] is an I(1) process

(2) If [[rho].sub.1] < 0, [[rho].sub.2] < 0 and (1 + [[rho].sub.1]) (1 + [[rho].sub.2]) <1, then the series is stationary and ergodic

(3) [H.sub.1] : [[rho].sub.1] < 0 and [[rho].sub.2] = 0

What if it is the intermediary case of a partial unit root ? Then,

[[rho].sub.1] < 0 and [[rho].sub.2] = 0

[H.sub.2] : or

[[rho].sub.1] = 0 and [[rho].sub.2] < 0

Here [y.sub.t] is stationary in one regime and nonstationary in another. Their test can distinguish amongst the three. The difficulty is that the null of a unit root ([[rho].sub.1] = [[rho].sub.2] = 0) is compatible with both the existence of a threshold ([[theta].sub.1] [not equal to] [[theta].sub.2]) or the nonexistence of a threshold ([[theta].sub.1] = [[theta].sub.2]). But CH determine that the assumptions of these two situations are different and hence we can simultaneously distinguish between nonstationarity and nonlinearity. Using theorems 5 and 6 of CH, the distinction between linearity and nonlinearity lies in the identification of the threshold parameter [lambda]. With no threshold effects, [lambda] is not identified, and so its estimate "[lambda][??]" is random (as always) and so is [R.sub.t] With threshold effects, [lambda] is identified and with no randomness in [R.sub.t], it is equivalent to the case where [[lambda].sub.0] is known. The unidentified threshold bootstrap imposes the restriction [theta] = [[theta].sub.1] = [[theta].sub.2] (no thresholds) and [rho] = 0 (unit root). In this case the bootstrap p-value is the percentage of simulated test statistic [R.sup.b.sub.t] that exceeds [R.sub.t]. The identified threshold bootstrap requires simulation of the TAR process, and calculating [R.sup.b.sub.t]. Again the bootstrap p-value is the percentage of simulated [R.sup.b.sub.t] that exceeds [R.sub.t].(4) Thus we conclude that in the presence of nonlinearity, the CH threshold unit root tests have more power than the standard ADF tests.

NONSTATIONARITY-NONLINEARITY TEST RESULTS

The results for Canada, Japan, United States and for the United Kingdom are in tables I through 4 respectively. First we test for the presence of threshold effects before rejecting a linear model in favor of a non-linear one.

From table 1, for Canada we see the Wald statistic [W.sub.t] for threshold variables of the form [Z.sub.t] = [y.sub.t] - [y.sub.tm] for delay parameters m = 1, ... 4 are significant across most lags (for m=1, 4 at the 5% level and for m=2 at the 10 per cent level) and therefore rejects the null hypothesis of linearity in favor of a threshold model. Since the results are sensitive to the choice of "m" it is necessary to select "m" endogenously. The least square estimate of m is equivalent to determining "m" such that [W.sub.T] is maximized. According to table 1, this corresponds to a value of m = 4. The p-value when m = 4 is significant at the 5% level. Even if the calculation of m is incorporated into the calculation of the bootstrap value, it is still significant at the 5% level, implying that it is likely that the GDP growth rate for Canada can be represented by a nonlinear model. Then we calculate the threshold unit root test statistic [R.sup.1.sub.t], [t.sub.1] and [t.sub.2] for all lags. All the bootstrap p-values of [R.sup.1.sub.t] are significant at the 1 per cent level. Caner and Hansen state that the [R.sub.1] statistic has power both against a stationary and a partial unit root alternative, and we should use the [t.sub.1] and [t.sub.2] statistic to determine which alternative hypothesis might apply. At m=4, the bootstrap p-value of [t.sub.1] = 0.0096 and of [t.sub.2] = 0.336, which is evidence in favor of [[rho].sub.1]<0 and [[rho].sub.2] = 0. This means that there is a unit root in the [y.sub.t] series only in the second regime (when change in the GDP growth rate is greater than 0.887). Therefore, the dynamics of the real GDP series growth rate in Canada are different in the two regimes.

From table 2, for Japan we see the Wald statistic [W.sub.t] for threshold variables of the form [Z.sub.t] = [y.sub.t] - [y.sub.tm] for delay parameters m = 1, ... 4 are significant only for m = 4 at the 10 per cent level and this gives weak support for rejecting the null hypothesis of linearity in favor of a threshold model. The endogenously determined value of m is 4 according to table 3. The p-value when m = 4 is significant at the 10% level. However, if the calculation of m is incorporated into the calculation of the bootstrap value, it is no longer significant even at the 10 per cent level, implying that it is highly unlikely that the GDP growth rate for Japan can be represented by a nonlinear model, and that a linear model is probably more appropriate.

From table 3, for the USA we see the Wald statistic [W.sub.t] for threshold variables of the form [Z.sub.t] = [y.sub.t] - [y.sub.t-m] for delay parameters m = 1, ... 4 are significant only for m = 1 and 4 at the 5 per cent level and this gives mixed support for rejecting the null hypothesis of linearity in favor of a threshold model. The endogenously determined value of m is 4 according to table 5 since the Wald statistic is maximized. The p-value when m=4 is significant at the 5% level. Even when the calculation of m is incorporated into the calculation of the bootstrap value, it is still significant at the 5 per cent level, implying that it is likely that the GDP growth rate for the United States can be represented by a nonlinear model. Then we calculate the threshold unit root test statistic [R.sup.1.sub.t], [t.sub.1] and [t.sub.2] for all lags. All the bootstrap p-values of [R.sup.1.sub.t] are significant at the 1 per cent level. At m = 4, the bootstrap p-value of [t.sub.1] = 0.245 and of [t.sub.2]=0.0001, which is evidence in favor of [[rho].sub.1]=0 and [[rho].sub.2]<0. This means that there is a unit root in the [y.sub.t] series only in the first regime (when change in the GDP growth rate is less than -3.55). Thus, the dynamics of the real GDP series growth rate in the United States are different in the two regimes.

From table 4, for UK we see the Wald statistic [w.sub.t] for threshold variables of the form [Z.sub.t] = [y.sub.1] - [y.sub.tm] for delay parameters m =1, ... 4 are significant only for m=2 and 3 at the 5 per cent level and for m = 1 at the 10 per cent level and this gives support for rejecting the null hypothesis of linearity in favor of a threshold model. The endogenously determined value of m is 2 according to table 7 since the Wald statistic is maximized. The p-value when m = 2 is significant at the 5% level. Even when the calculation of m is incorporated into the calculation of the bootstrap value, it is still significant at the 5 per cent level, implying that it is likely that the GDP growth rate for the United Kingdom can be represented by a nonlinear model. Then we calculate the threshold unit root test statistic [R.sup.1.sub.t], [t.sub.1] and [t.sub.2] for all lags. All the bootstrap p-values of [R.sup.1.sub.t] are significant at the 1 per cent level. At m = 4, the bootstrap p-value of [t.sub.1] = 0.0.0083 and of [t.sub.2] = 0.112, which is evidence in favor of [[rho].sub.1] <0 and [[rho].sub.2]=0. This means that there is a unit root in the [y.sub.t] series only in the second regime (when change in the GDP growth rate is greater than -4.61). This means that the dynamics of the real GDP series growth rate in the United Kingdom are also different in the two regimes.

Thus our results indicate nonstationarity in the GDP growth rates for Canada, United States and the United Kingdom, and stationarity in the GDP growth rate for Japan. This implies that there are regime changes in Canada, United States and United Kingdom, and no regime changes in Japan. This is perhaps due to the fact that in Canada, U.S.A and U.K there have been changes in the party in power at the center over the sample period, whereas in Japan, except for a brief period, the same party has been in power over the entire sample period. The lack on any non-linearity in Japan can be explained, at least partially, by the anemic growth rates of GDP in the 1990s, which obviously imply that there weren't any major structural changes. It has taken the Japanese government a long time to deal with the structural problems of the Japanese economy, and while the previous policies were in place, there were no structural changes and therefore no non-linearities.

Moreover, for Canada and the United Kingdom we find that the first regime (GDP growth rate is less than the threshold) has a unit root (is mean reverting) and the second regime (GDP growth rate is greater than the threshold) is stationary, whereas for the United States we find that the second regime has a unit root and the first regime is stationary. This implies that for Canada and the U.K. the government intervenes when the changes in the GDP growth rates are small, and the government does not intervene when there are large changes in the growth rate. This would imply that the government's objective is to stimulate the economy, and inflation is not a target. However, for the U.S. the government / Federal Reserve intervenes when the changes in the GDP growth rate are high, implying that inflation possibly is a target. For the U.S. the stationary first regime implies no intervention when changes are small.

An alternate explanation for this is the Romer (1990) growth model, who states that "..technological change ... lies at the hearth of economic growth," and "technological change arises in large part because of intentional actions taken by people who respond to market incentives." (5) An implication of this model is that "an economy with a larger stock of human capital will experience faster economic growth." (6) The faster growth rates of GDP in the United States compared to Canada, U.K. and Japan in the 1990s can be explained by the rapid rate of advancement of technology in the U.S. compared to the other countries. This growth of technology can, at least partially, be explained by the fact that throughout this period the U.S. was the preferred destination for students and skilled workers throughout the world, and this increase in human capital enabled technology to improve and act as a prime accelerant for GDP growth in the U.S. Given the larger role played by the private sector in the United States economy, market incentives for technological progress were more significant in the U.S. compared to the other three countries in our sample.

This would also explain the non-linearities in our results. If technology was progressing more rapidly in the United States, perhaps due to market incentives being more significant, the GDP growth rate would also be more rapid and non-linear, since market incentives were not likely to be spread evenly across the economy and over time. When incentives are the greatest, GDP growth rate would also be the greatest as the U.S. economy is driven by the private sector. If the government is concerned about inflation, this would give the government/Federal Reserve a reason to intervene to keep inflation under control. Therefore, in the United States we have an unit root in the second regime (GDP growth rate is greater than the threshold), whereas the first regime is stationary as when GDP growth rates are low the private sector does not have an incentive to invest in research and development and improve technology as the potential returns from doing so are low. In Canada and the United Kingdom the government plays a much bigger role in the economy, and consequently the private sector plays a smaller role. These two countries also have access to less human capital (for a variety of reasons) compared to the United States. Therefore, the pace of technological improvement will be less than in the United States. Due to the smaller size of the private sector, when GDP growth rates are small, the government often intervenes to attempt to stimulate research and development and therefore technological change. This would lead to non-linearities in the GDP growth rates in the first regime in these two countries. When GDP growth rates improve, and the government steps back, the smaller size of the private sector (and of human capital) in these two countries would mean that innovation and growth will not fluctuate a great deal, and are likely to be linear.

CONCLUSION

To summarize our results, we find significant evidence indicating that the real GDP growth rate in Canada, U.S.A. and U.K. is non-linear, while for Japan it is linear. An implication is that in Japan policymakers react in a similar way without regard to whether the changes in real GDP growth rate are large or small. This may be due to the fact that the same political party has been in power in Japan (with a brief exception) over the entire sample period. Also, the anemic growth rates of GDP in Japan in the 1990s haven't given the private sector much incentive to invest in technological progress, leading to linear changes. Canada, U.S.A. and U.K. have experienced changes in government and therefore changes in policy over this period, and hence have a nonlinear response to changes in the real GDP growth rate. Policymakers in Canada and the U.K. seem to react when changes are below the threshold (small changes). This could be because for political reasons they are targeting growth and not inflation. The larger role played by the government and the consequent smaller role for the private sector (due at least in part to the more limited access to human capital) implies that at smaller growth rates of GDP there is more government intervention, which together with the private sector might lead to non-linear changes in GDP growth rate. At higher growth rates of GDP when the private sector is left alone by the government, the lack of human capital might lead to more gradual changes in GDP. On the other hand, in the United States the access to human capital and the significantly larger private sector means that technology will improve most rapidly at higher growth rates when the private sector has an incentive to invest in research and development, but at lower growth rates the incentives are absent, and therefore the private sector will reduce investment, and the government's role in the economy is not large enough to have a significant impact. This also implies that the U.S. government/Federal Reserve will be primarily concerned with targeting inflation as they know that there is not much they can do with regards to economic growth, which is determined by the private sector. In Canada and the United Kingdom, the governments will be concerned with economic growth as they know that the private sector is unwilling or incapable of stimulating the economy by themselves.

[GRAPHIC OMITTED]

ACKNOWLEDGEMENT

The authors would like to thank Bruce Hansen for making the GAUSS programs available. Dutt would like to thank the State University of West Georgia for their Faculty research grant (#: 10000-1014408-12100-11000: 2004-05).

REFERENCES

Ahmed, S, Levin, A and Wilson, B. A. (2002)," Recent US Macroeconomic Stability: Good Policy, Good Practices or Good Luck?" International Finance Discussion Paper # 730, Board of Governors of the Federal Reserve System.

Ahn, D and Gao, B (1999): "A Parametric Nonlinear Model of Term Structure Dynamics," Review of Financial Studies, 12, 721-762.

Beaudry, P. and Koop, G. (1993), "Do Recessions Permanently Change Output?" Journal of Monetary Economics, 31(2), pp. 149-63.

Balke, N. S. and Wynne, M. A. (1995), "Recessions and Recoveries in Real Business Cycle Models," Economic Inquiry, 33, pp. 640-63.

Bradley, M. D. and Jansen, D. W. (1997), "Nonlinear Business Cycle Dynamics: Cross Country Evidence on the Persistence of Aggregate Shocks," Economic Inquiry, 35, pp. 495-509.

Breunig, R. and Stegman, A. (2003), Testing for Regime Switching in Singaporean Business Cycles, Working Paper, Australian National University, Australia.

Bodaman, P. and Crosby, M. (1999), "Nonlinearities in the Singaporean Business Cycle," Singapore Economic Review, 44(2), pp. 24-35.

Blanchard, O. and Simon, J. (2001), "The Long and Large Decline in US Output Volatility," Brooking Papers on Economic Activity, 2001(1), pp. 135-64.

Box, G. and Jenkins, G. (1976), Time Series Analysis: Forecasting and Control, Holden-Day: San Francisco.

Braun, P. and Zarnowitz, V. (1989), Major Macroeconomic Variables and Leading Indexes: Some Estimates of their Interrelations, 1886-1982, NBER Working paper # 2812.

Caner, M. and Hansen, B. E. (2001), "Threshold Autoregression With a Unit Root," Econometrica, 69, 1555-1596.

Chauvet, M and Potter, S. (2001), "Recent Changes in US Business Cycle," The Manchester School, 69, pp. 481-508.

Cheung, Y. W. and Chinn, M. D. (1996), "Deterministic, Stochastic and Segmented Trends in Aggregate Output: A Cross Country Analysis," Oxford Economic Papers, 48, pp. 134-62.

Darbha, G. and Patel, U. R. (2004), Nonlinear Adjustment in Real Exchange Rates and Long Run Purchasing Power Parity-Further Evidence, Working Paper, National Stock Exchange, India.

DeJong, D. N., Dharmarajan, H., Liesenfeld, R. and Richard, J. F., (2004), On the Structural Stability of US GDP, Working Paper, University of Pittsburgh.

DeLong, J. B. and Summers, L. H. (1986), "Are Business Cycles Symmetrical," in American Business Cycle: Continuity and Change, ed. R. J. Gordon, The University of Chicago Press, pp. 166-79.

Diebold, F. and Rudebusch, A. (1991), "Forecasting Output with the Composite Leading Index: A Real Time Analysis," Journal of the American Statistical Association, 86, pp. 603-610.

Durland, J. M. and McCurdy, T. H. (1994), "Duration Dependent Transitions in a Markov Model of US GNP Growth," Journal of Business and Economic Statistics, 12, pp. 279-88.

Diebold, F. X. and Redebusch, G. (1990), "A Nonparametric Investigation of Duration Dependence in the American Business Cycle, Journal of Political Economy, 98, pp. 596-616.

Enders, W. and Granger, C. W. J. (1998), "Unit Root Tests and Asymmetric Adjustment with an Example Using the Term Structure of Interest Rates," Journal of Business and Economic Statistics, 16, pp. 304-12.

Falk, B. (1986), "Further Evidence on the Asymmetric Behavior of Economic Time Series Over the Business Cycle," Journal of Political Economy, 94, pp. 1096-1109.

Filardo, A. J. and Gordon, S. F. (1998), "Business Cycle Durations," Journal of Econometrics, 85, pp. 99-123.

Galvao, A. B. (2002), Can Nonlinear Time Series Models Generate US Business Cycle Asymmetric Shape? Working Paper, European University Institute.

Gonzalez, M. and J. Gonzalo (1998), Threshold Unit root Models, Working Paper, U. Carlos II de Madrid.

Hamilton, J. D. (1989), "A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle," Econometrica, 57, pp. 357-84.

Harding, D. and Pagan, A. R. (2002), "Dissecting the Cycle: a Methodological Investigation," Journal of Monetary Economics, 49, pp. 365-81.

Jones, C. S. (2003), "Nonlinear Mean Reversion in the Short Term Interest Rate," The Review of Financial Studies, 16(3), 793-843.

Keynes, J. M. (1936), The General Theory of Employment, Interest and Money, London, MacMillan.

Kim, C. and Nelson, C. R. (1998), "Business Cycle Turning Points; A New Coincident Index and Tests of Duration Dependence Based on a Dynamic Factor Model with Regime Switching," Review of Economics and Statistics, 80, pp. 188-201.

Kim, C. J. and Nelson, C. R. (1999 a), "Has the US Economy Become More Stable? A Bayesian Approach on a Marker Switching Model of the Business Cycle," Review of Economics and Statistics, 81, pp. 608-616.

--(1999 b), "Friedman's Plucking Model of Business Fluctuations: Tests and Estimates of Permanent and Transitory Components," Journal of Money, Credit and Banking, 31(3), pp. 317-34.

Kim, C., Morley, J. and Piger, J. (2004), "Markov Switching Model of Business Cycle Dynamics with a Post Recession Bounce Effect," Journal of Business and Economic Statistics, 22(1), pp. 80-93.

Lam, P. (1990), "The Hamiltonian Model with a General Autoregressive Component; Estimation and Comparison with other Models of Economic Time Series," Journal of Monetary Economics, 26, pp. 409-32.

--(2004), "A Markov Switching Model of GNP Growth with Duration Dependence," International Economic Review, 45, pp. 175-204.

McConnell, M. M. and Perez-Quiros, G. (2000), "Output Fluctuations in the United States: What has Changed Since the Early 1980s?," American Economic Review, 90 (5), pp. 1464-76.

McQueen, G. and Thorley, S. (1993), "Asymmetric Business Cycle Turning Points," Journal of Monetary Economics, 31, pp. 341-62.

Neftci, S. N. (1984), "Are Economic Time Series Asymmetric over the Business Cycle," Journal of Political Economy, 92, pp. 307-28.

Nelson, C. R. and Plosser, C. I. (1982), "Trends and Random Walks in Macroeconomic Time Series: Some Evidence and Implication," Journal of Monetary Economics, 10, pp. 139-62.

Niemira, M. P. and Klien, P. A. (1994), Forecasting Financial and Economic Cycles, New York, John Wiley and Sons, Inc.

Pesaran, M. H. and Potter, S. M. (1997), "A Floor and Ceiling Model of US Output," Journal of Economic Dynamics and Control, 21, pp. 661-95.

Rapach, D. E. (2002), "Are Real GDP Levels Stationary? Evidence from Panel Data Sets," Southern Economic Journal, 68, pp. 473-95.

Romer, Paul M. (1990), "Endogenous Technological Change," Journal of Political Economy, Vol. 98, No. 5, S71-S102.

Sichel, D. E. (1993), "Business Cycle Asymmetry: A Deeper Look," Economic Inquiry, 31, pp. 224-36.

--(1994), "Inventories and the Three Phases of the Business Cycle," Journal of Business and Economic Statistics, 12, pp. 270-77.

Stock, J. H. and Watson, M. W. (2002), "Has the Business Cycle Changed and why," NBER Macroeconomics Annual, 17, pp. 159-218.

Tsay, R. S. (1997), "Unit Root Tests with Threshold Innovations," University of Chicago.

Watson, M. W. (1986), "Univariate Detrending Methods with Stochastic Trends," Journal of Monetary Economics, 18, pp. 29-75.

Wasserfallen, W. (1986), "Nonstationarities in Macroeconomic Time Series-Further Evidence and Implication," Canadian Journal of Economics, 19, pp. 498-519.

Warnock, M. V. and Warnock, F. F. (2000), "The Declining Volatility of US Employment : was Arthur Burns Right?" International Finance Discussion Paper # 677, Board of Governors of the Federal Reserve System.

SWARNA (BASU) DUTT

University of West Georgia, Carrollton

DIPAK GHOSH

Emporia State University, Kansas

ADRIAN M. AUSTIN

University of West Georgia, Carrollton

NOTES

(1.) Though some nonlinear setups cannot replicate the asymmetries well, some regime switching models with changing variances perform much better than constant variance models.

(2.) This paragraph is heavily dependent on Darbha and Patel (2004).

(3.) Tsay (1997) introduces unit root tests in the presence of threshold effects, but the autoregressive lags are constant across regimes (which is not true here) making it a special case of the CH methodology. Also, Gonzalez and Gonzalo (1998) examine a TAR(l) model with nonstationarity, but of a particular type, namely geometrically ergodic.

(4.) CH run Monte Carlo simulations to show their relative strength vis-a-vis the conventional Dickey-Fuller (ADF) tests in the presence of thresholds.

Case 1: The condition [rho]1 = [rho]2 is imposed and [DELTA][mu] = 0 (no regimes), the ADF is more powerful than the CH threshold unit root test. But as [DELTA][mu] increases, the [R.sub.1t] and [R.sub.2t] tests are more powerful than the ADF test.

Case 2: This is where [rho]1 = 0, [rho]2 varies and [DELTA][mu] = 0, a partial unit root model. Here [R.sub.1t] and [R.sub.2t] have substantially greater power than the ADF test. The ADF test is particularly weak when [DELTA][mu] is large. Here the t-ratio test is itself enough to distinguish between the pure unit root, the partial unit root and the stationary cases.

Case 3: [[rho].sub.1] is fixed, [[rho].sub.2] varies and [DELTA][mu] = 0, the stationary case. Here [R.sub.1t] is the most powerful test with [R.sub.2t] a close second.

(5,6.) Romer (1990, pp. S72)

(6.) Romer (1990, pp S99)

Table 1: Canada
Threshold and Unit Root Tests Unconstrained Model

Bootstrap Threshold Test                  Unit Root Tests, p-Value

                                       [R.sub.1T]

M    [W.sub.T]   1%C.V.    p-Value       Asym       Boot

1      21.7       26.4       0.0400       0.0055   0.0094
2      19.5       26.6       0.0795       0.0022   0.0048
3      16.7       27.1        0.162       0.0003   0.0016
4      25.5       27.0       0.0135       0.0054   0.0096

Bootstrap Threshold Test

               [t.sub.1]                [t.sub.2]

M                 Asym          Boot       Asym         Boot

1                 0.961        0.764      0.0023       0.0029
2                 0.742        0.384      0.0021       0.0031
3                 0.799        0.447      0.0002       0.0005
4                0.0076       0.0052       0.614        0.336

Notes for tables 1, 2, 3 and 4:

(1) Bootstrap p-values are calculated from 10,000 replications

(2) "M" is the delay parameter

(3) [W.sub.T] is the Wald statistic which tests the null hypothesis of
linearity against the alternate hypothesis of a threshold effect.

(4) The [R.sub.1] statistic tests the null hypothesis of a unit root
against the alternative hypothesis that either the first or the second
regime is threshold stationary (the asymptotic and bootstrap p- values
for the test statistic are reported.

(5) The [t.sub.1] and [t.sub.2] test the null hypothesis of an unit
root against the alternate hypothesis of stationarity in the first and
the second regime respectively (the asymptotic and bootstrap p-values
are reported)

Table 2 : Japan
Threshold and Unit Root Tests Unconstrained Model

Bootstrap Threshold Test                        Unit Root Tests, p-Value

                                       [R.sub.1T]

M    [W.sub.T]   1%C.V.    p-Value       Asym       Boot

1      16.1       27.8        0.188       0.0958   0.0754
2      11.5       27.2        0.530       0.0140   0.0186
3      8.88       27.7        0.783        0.181    0.135
4      18.8       27.5       0.0908       0.0053   0.0102

Bootstrap Threshold Test

       [t.sub.1]               [t.sub.2]

M         Asym         Boot      Asym         Boot

1         0.768       0.959      0.0448       0.0290
2         0.813       0.948      0.0059       0.0063
3         0.508       0.248       0.311        0.151
4         0.460       0.989      0.0022       0.0022

Table 3 : USA
Threshold and Unit Root Tests Unconstrained Model

Bootstrap Threshold Test                Unit Root Tests, p-Value

                                        [R.sub.1T]

M     [W.sub.T]    1%C.V.     p-Value      Asym       Boot

1       21.4        25.4       0.0396     0.0000      0.000
2       11.1        26.4        0.522     0.0000     0.0002
3       13.5        25.4        0.312     0.0000      0.000
4       26.1        25.6       0.0089     0.0000     0.0001

Bootstrap Threshold Test

        [t.sub.1]                [t.sub.2]

M          Asym         Boot       Asym         Boot

1         0.0010       0.0020     0.0008       0.0008
2          0.223        0.119     0.0002       0.0006
3         0.0379       0.0282     0.0002       0.0005
4         0.0456        0.245     0.0000       0.0001

Table 4 : UK Threshold and Unit Root Tests Unconstrained Model

Bootstrap Threshold Test             Unit Root Tests, p-Value

                                     [R.sub.1T]

M     [W.sub.T]   1%C.V.   p-Value     Asym       Boot

1       18.9       27.0     0.0935    0.0000     0.0011
2       43.3       27.3     0.0004    0.0003     0.0018
3       28.0       27.4     0.0101    0.0002     0.0015
4       12.0       26.9      0.427    0.0010     0.0043

Bootstrap Threshold Test

      [t.sub.1]              [t.sub.2]

M        Asym        Boot     Asym        Boot

1       0.0986      0.0547    0.0003      0.0021
2       0.0642       0.404    0.0098      0.0126
3        0.957       0.743    0.0000      0.0005
4       0.0074      0.0083     0.212       0.112