Oil Price Shocks and Real GDP Growth Empirical Evidence for Some OECD Countries

Transcription

1 Oil Price Shocks and Real GDP Growth Empirical Evidence for Some OECD Countries Rebeca JiménezRodríguez University of Alicante Marcelo Sánchez European Central Bank First version: November 2002 This version: April 2003 y Abstract Most of the studies existing in theoretical and empirical understanding of the macroeconomic consequences ofoil price shocks have been focused on the US economy. In contrast these studies, this paper assesses empirically the e ects ofoil price shockson the real economic activity ofthe main industrialised OECD countries (individual G7 countries, Norway and the euro area as a whole), distinguishing between oil importing and oil exporting countries. We use multivariate VAR analysis for each country, considering both linear and nonlinear approaches to modelling the role of oil prices. The pattern of responses to oil price shocks by GDP growth is broadly similar across oil importing countries, with oil price increases having a negative impact on GDP growth. With regard to oil exporters, the e ect of oil shocks on GDP growth di ers between the UK and Norway, with the rst country being negatively a ected by an oil price increase and the latter bene ting from it. In two nonlinear speci cations that allow us to distinguish between the impact of oil price rises and falls, we observe that oil price declines have a signi cant e ect only in two or three countries, depending on the speci cation. Furthermore, we provide some evidence suggesting that nonlinearmodelling of the relationshipbetween oil pricesandreal activity appearsto performbetterthana linearapproach. Keywords: Macroeconomic uctuations; Oil price shock; Nonlinear speci cations. JEL codes: E32, Q43. The views expressed in this paper are those of the authors and do not necessarily re ect the position of the European Central Bank. y Correspondence to: Departamento de Fundamentos del Análisis Económico. University of Alicante Alicante. Spain. Tel.: Ext.: Fax.: rebecaj@merlin.fae.ua.es; or European Central Bank. Kaiserstrasse Frankfurt. Germany. Tel.: Fax.: marcelo.sanchez@ecb.int 1

2 1 Introduction A large body of research suggests that oil price uctuations have considerable macroeconomic consequences. These consequences on real activity are expected to be di erent in oil importing and in oil exportingcountries. Whereas an oil price increase should be considered good news in oil exporting countries and bad news in oil importing countries, the reverse should be expected when oil price decreases. The di erent transmission mechanisms of oil price shocks the real economy have di erent features in each of the countries, which thus respond somewhat di erently to such shocks in light of the di erent levels of oildependence, the di erent policies implemented to smooth out the consequences of such shocks, etc. Finally, the literature seems to have reached a consensus that the impact of oil prices on macroeconomic variables need not to be linear, but may rather have asymmetric e ects or adopt other type of nonlinear behaviour. Against this background, this paper, in contrast to most of the studies existingin the literaturewhich are focused on the US economy, empirically assesses the e ects of oil price shocks 1 considering both linear and nonlinear approaches in the main industrialised OECD countries, taking both oil importing and exporting countries into account. The channels through which oil prices have an impact on economic activity include those that a ect thesupply and demand sides of the economy. The supply side e ects involve both those directly on rms and those on the labour market. With regard the e ects on rms, given that crude oil is a basic input to production, an increase in oil price leads to a rise in production costs that induces companies to lower output. 2 With regard the overall e ect of an increase in oil prices on the labour market, it is hard to assess. On the one hand, the increase in the price of a substitute input such as oil implies a higher demand for labour and higher real wages (particularly if capital is complementary to oil), while, on the other hand, the lower level of production induced by the same shock exerts a downward pressure on the demand for labour and thus tends to reduce real wages. In addition the e ects on supply, higher oil prices entail demandside e ects on consumption and investment. Consumption will decrease in an oil exporting country when oil prices decline, and in an oil importing country when oil prices increase. 3 The more intensive oil shocks are these impacts on consumption will be stronger and thus changes in real disposable income are expected to be persistent. If consumers believe instead that the increase in oil prices is temporary, they will attempt to smooth out their consumption by saving less or borrowing 1 We consider four crises: ArabIsraeli War (1973), Iranian Revolution (1978), IranIraq War (1980), and Persian Gulf War (1990). 2 See, for example, Rasche and Tatom (1977, 1981), and Kim and Loungani (1992) who introduced the increase in oil price as a standard supplyside shock that reduces potential output. 3 It has normally been the case that international spillovers through trade have reinforced these e ects among oilimporting industrial economies. This is due the fact that industrial economies trade mostly among themselves (which are overall negatively a ected by oil prices) and that oil exporting countries do not use all the extra income induced by high oil prices in ways that end up increasing their demand for foreign products. 2

3 more, with real consumption little a ected. On top of the e ects on consumption, oil prices will have a negative impact on investment through their impact on rms revenues, also particularly if the oil shock is expected to be relatively persistent. The explanation is that, at higher oil prices, the demand for capital (in the absence of a strong substitution e ect in production) will drop, thereby inducing a downward adjustment in investment activity. Finally, in addition the previously discussed impacts of oil prices on supply and demand, oil price changes normally impact foreign exchange markets and in ation, giving thus rise to indirect e ects on real activity. Indeed, real exchange rates tend to move to partly o set the termsoftrade impact of oil prices on real output, depreciating when oil prices induce a positive termsoftrade shock and appreciating when oil prices entail a negative termsoftrade shock. Moreover, oil shocks are in ationary, makingsome centralbanks adopt atighter monetary policy, therebycontributing to induce a slowdown in economic activity. 4 Evidence of a negative correlation between oil price movements and macroeconomic uctuations in oil importing countries have been found since the 1980s, among others, by Rasche and Tatom (1981), Darby (1982), Hamilton (1983), Burbidge and Harrison (1984), and Gisser and Goodwin (1986). 5 Some of them also reported that oil price Grangercauses output variables. Both the negative correlation and the Granger causality results, however, were later found to be weaker when datafrom the mid1980s are included. In fact, the declines in oil prices occurred over the second half of the 1980s had smaller positive e ects on economic activity. 6 Thus, Mork (1989), 7 Lee et al. (1995) and Hamilton (1996) introduced nonlinear transformations of oil price to reestablish the negative correlation between increases in oil prices and economic downturns, as well as the Granger causality. 8 More recently, JimenezRodriguez (2002a, 2002b) and Hamilton (2003) also found evidence of a nonlinear relationship between the two variables for the US economy. While the introduction of nonlinear speci cations to characterise the relationship between oil prices and real activity was motivated by the fact that the linear relationship seemed to weaken 9 over time, 4 Some authors, like Bohi (1989), and Bernanke et al. (1997), argue that economic downturns observed in the US after oil price shocks are caused by a combination of direct impacts of the shocks themselves and the monetary responses them. 5 While all these studies analyses this relationship for the US, Darby (1982) and Burbidge and Harrison (1984) also considered other developed countries (Japan, Germany, the UK, Canada, France, Italy, and the Netherlands in the former case, and Japan, Germany, the UK and Canada in the latter). 6 A good example is the fact that an acceleration of U.S. economic activity did not follow the sizeable oil price decline that occurred since late Mork s study found that the e ects of oil price increases are di erent from those of decreases, and that oil price decreases are not statistically signi cant in the US. This study has proved in uential in that many authors have thereafter not even considered the possibilityof e ects derived from adecrease in oil prices. 8 Mork et al. (1994) documented the asymmetry in the inverse relationship between oil price and aggregate economic activity for countries other than the US (includingboth oilimporting and oilexporting countries). In particular, they found asymmetry in the cases of Norway and all G7 countries but Italy. 9 It is worth noting that Bacon (1991), Karrenbock (1991), Balke et al. (1998), all also found evidence of nonlinearity, observing an asymmetric response in gasoline market. The common idea is that gasoline prices rise more quickly when oil 3

4 an economic justi cation for the asymmetry was also o ered in the literature. This explanation relates to adjustment costs resulting from the implied sectoral reallocation of resources. For instance, Lilien (1982) has formulated the socalled dispersion hypothesis, arguing that a change in oil price alters the equilibrium relation between various sectors. 10 For example, an increase (decrease) in oil prices would lead to a contraction (expansion) in sectors that make use of oil in the production process. Moreover, such increase (decrease) in oil prices would generate an expansion (contraction) of energye cient sectors relative to energyintensive sectors. However, given that in the short run the cost of reallocation of resources between sectors is high, oil shocks that imply readjustment between energye cient and energyintensive sectors will give rise to some overall loss in output. While this loss will aggravate the economic contraction when oil prices increase, it will constrain the economic expansion when oil prices decline, thereby giving rise the asymmetric e ect. Based on the previous empirical literature and economic arguments, this paper assesses empirically the possibility of asymmetry in the e ect of oil prices on real activity. Moreover, given that the asymmetry is a very special case of nonlinear relationship between GDP and oil prices, in this paper we consider two other nonlinear transformations proposed in the literature, namely: scaled speci cation (Lee et al., 1995), taking the volatility of oil prices into account; and net speci cation (Hamilton, 1996), considering the net amount by which oil prices have gone up over the last year. 11 Unlike most of the studies existing in the literature, which focus on the United States economy, 12 we analyse the e ects of an oil price shock in the main industrialised countries (individual G7 countries, Norway and the euro area as a whole). While most of these countries are oil importing, we also include in our sample two net oilexporting countries, namely the UK and Norway (see Figure 1). 13 The main objective of this paper is to analyse the e ects of oil price shocks on the economic activity of the main industrialised OECD countries, distinguishing between oil importing and oil exporting countries. In doingso, we use both linear and nonlinear approaches tomodellingtherelationship between oilprices and GDP growth. To our knowledge, this paper is the rst to assess the impact of oil price shocks on the level of real activity in the main industrialised OECD countries considering di erent linear and nonlinear speci cations. prices are increasing than they fall when oil price are decreasing. 10 In the same line of thought are, among others, Loungani (1986), Davis (1987), Hamilton (1988), Davis et al. (1997). 11 A more comprehensive analysis, which would include consideration of kernel semiparametric speci cation (see Jimenez Rodriguez, 2002b, for the case of the US) was not possible as it would require samples longer than available for most countries under study. 12 Some of the few relevant exceptions are cited in footnotes 5 and Norway and the UK switched from a position of net oil importing to oil exporting in the 1970s. Canada also switched its position at the beginning of the 1980s (See Figure 1), making it hard to interpret the results for Canada. Since our sample starts in 1972:III, Canada has been a net oil importer during three of the four oil crises considered. As such, we thereafter consider Canada tentatively as a net oil importing country. 4

5 Our main ndings may be summarised as follows. Preliminary Grangercausality type analysis permits us to conclude that the interaction between oil variables and macroeconomic variables is found to be signi cant, with the direction of causality going in at least one direction in all countries and in both directions in most countries. The e ects of an increase in oil prices on real economy and those of a decrease di er substantially, showing evidence against the linear approach that assumes that oil price declines are as bene cial as oil price increases of the same magnitude are detrimental to real economic activity. An increase in oil prices, on the one hand, is found to have a signi cant negative impact on the GDP growth in all oil importing countries but Japan and in oil exporter UK (relating the standard Dutch disease e ect), while it a ects positively Norwegian GDP growth. In addition, the e ects of an oil price hike on GDP growth are stronger for the US, although euro area countries (Germany, France and Italy) exhibit similar strong real e ects when we use the nonlinear approach. On the other hand, in the asymmetric and scaled speci cations, which allow us to distinguish between oil price rises and falls, we nd that a decline in oil prices a ects signi cantly only few countries, having a positive impact on the US and UK economies while doingharm tothecanadian economy. Evidenceon therelative performance of the di erent speci cations suggests that the speci cation that considers the volatility of oil prices (i.e. thescaled speci cation) performs in all countries somewhat than the other ones. Additionally, we do not nd evidence of instability of the results for the GDP equation with the exception of the linear model for the euro area. The paper is organised as follows. Section 2 describes the methodology. Section 3 presents the empirical results. Concluding remarks are o ered in Section 4. 2 Methodology We consider the following vector autoregression model of order p (or simply, VAR(p)) 14 : px B 0 y t = k B i y t i u t ; (1) i=1 where y t is a (n 1) vector of endogenous variables, B 0 and B i are (n n) matrices of coe cients, k is a vector of constants, p is the number of lags, and u t is a (n 1) of uncorrelated white noise disturbances. The matrix B 0 is taken to be lower triangular with 1 s along diagonal. This assumption guarantees that the VAR(p) model is just identi ed. We therefore can rewrite it as: 14 Sims (1980) was the rst to represent the reduced form of a standard openeconomy macroeconomic model as a multivariate dynamic system. This dynamic model was given by an unconstrained (i.e. restrictions only on lag length) vector autoregression. 5

6 y t = c px i y t i " t, (2) i=1 where c = B 1 0 k, i = B 1 0 B i for i = 1;::;p, " t = B 1 0 u t is a (n 1) vector of white noise disturbances. We consider a quarterly sevenvariable VAR for each country under study. The variables considered for the model are the following: real GDP, the real e ective exchange rate (REER), 15 real oil price, real wage, in ation, and short and longterm interest rates. Some variables (real GDP, REER, real oil price and real wage) are expressed in logs, while the remaining ones are simply de ned in levels. We include real oil prices and real GDP growth since our main objective is to analyse the e ects of oil shocks on the GDP growth. The remaining variables are included to capture some of the most important transmission channels through which oil prices may a ect economic activity indirectly. Before studying the e ects of oil shocks on real economic activity, we investigate the stochastic properties of the series considered in the model by analysing their order of integration on the basis of a series of unit root tests. Speci cally, we perform the DFGLS and P T tests of Elliott et al. (1996), and the DFGLSu and Q T tests of Elliott (1999), as well as the Augmented DickeyFuller (ADF) test (See Appendix 1). 16 Results of these formal tests are summarised in Tables 1A1C, indicating that rst di erence of all seven variables are stationary. We therefore de ne the vector y t in (2) to be given by the rst logdi erences of the rst four aforementioned variables (realgdp, REER, realoil price, and realwage), alongwith the rst di erences of theremaining ones (in ation, and short and longterm interest rates). The VAR in (2) can be estimated under the assumption that " t is a Gaussian white noise process by Maximum Likelihood, which is equivalent to estimating the model by OLS equation by equation. Moreover, even in case the true innovations were nongaussian, the OLS estimates thus obtained are consistent. The VAR system can be transformed into its Moving Average representation in order to analyse the system s response to a real oil price shock y t = ¹ 1X ª i " t i, (3) P where ª 0 is the identity matrix, and ¹ is the mean of the process (¹ = (I n p i ) 1 c). i=0 In order to assess the impact of shocks on endogenous variables, we examine the orthogonalised impulseresponse functions, using a Cholesky decomposition, as well as the accumulated responses. To 15 REER is de ned such that an increase means a real appreciation of the currency considered. An appreciation of the real exchange rate is expected to hurt the country s external competitiveness. 16 It is worth noting that the tests proposed by Elliott et al. (1996) and Elliott (1999) are more powerful than the DickeyFuller test. i=1 6

7 do so, we should choose an ordering for the variables in the system, since this method of orthogonalisation involves the assignment of contemporaneous correlation only to speci c series. Thus, the rst variable in the ordering is not contemporaneously a ected by shocks the remaining variables, but shocks the rst variable do a ect the other variables in the system; the second variable a ects contemporaneously the other variables (with the exception of the rst one), but it is not contemporaneously a ected by them; and so on. In our case, we have assumed the following ordering 17 : real GDP, real oil price, in ation, shortterm interest rate, longterm interest rate, real wage, and REER. The two standard error bands around the impulse responses are based on Lutkepohl (1990). We start by estimating a linear speci cation of the VAR in (2). Furthermore, on the basis of the previous empirical literature and economic arguments we also consider three nonlinear transformations of oil prices. Such nonlinear transformations are the following: 1) asymmetric speci cation, in which increases and decreases in the price of oil are considered as separate variables; 2) scaled speci cation (Lee et al., 1995), which takes the volatility of oil prices into account; and 3) net speci cation (Hamilton, 1996), wheretherelevant oilpricevariableis characterised bytheamount bywhich theseprices in quarter t exceed the maximum value over the previous four quarters. 18 The asymmetric speci cation distinguishes between the positive rate of change in the oil price, o t, and its negative rate of change, o t, which are de ned as follows: 8 < o if > 0 t = : 0 otherwise 8 < o t = : if < 0 0 otherwise where is the rate of change in the real oil price. The scaled and net speci cations were developed by Lee et al. (1995) and Hamilton (1996), respectively, to account for the fact that oil price increases after a long period of price stability had more dramatic macroeconomic consequences than those that were merely corrections to greater oil price decreases during the previous quarter. Lee et al. (1995) thus proposed the following AR(4)GARCH(1,1) 17 The ordering considered implies that we do not allow for the contemporaneous in uence of real oil price innovation over GDP growth. However, we have veri ed that the impulse responses do not change considerably when we consider the contemporaneous e ects of an oil price innovation on real GDP growth by placing the former at the top of the list of variables. 18 Unlike Hamilton, who uses nominal oil price, Mork and Lee et al. use the real price of oil. The present paper follows the latter approach. 7

8 representation of oil prices: = e t e t ji t 1» N(0;h t ) h t = 0 1 e 2 t 1 2 h t 1 q SOPI t = max(0; ^e t = ^h t ) q SOPD t = min(0;^e t = ^h t ) where SOPI stands for scaled oil price increases, while SOPD for scaled oil price decreases. Hamilton (1996) proposed a di erent nonlinear transformation, by using as an explanatory variable what he calls net oil price increase (NOPI). This variable is de ned to be the amount by which (the log of) oil prices in quarter t, p t, exceed the maximum value over the previous 4 quarters; and 0 otherwise. That is: NOP I t = maxf0;p t maxfp t 1 ;p t 2 ;p t 3 ;p t 4 gg The sample period considered is common to all countries under study, and it runs (including the lagged initial values) from 1972:III to 2001:IV, for a total of T = 118 available quarterly observations (See Appendix 2). To nd the suitable lag length, we perform di erent tests, namely, the Sims (1980) modi cation of the Likelihood Ratiest, as well as the Akaike, Schwartz and HannaQuinn Criteria. Whenever there is disagreement among the di erent tests, the optimal lag length is chosen using the Likelihood Ratiest. 3 Empirical Results In this section we analysethe empirical results for the four di erent oil price speci cations described in the previous section. In subsection 3.1 we test for the signi cance of the di erent oil measures and carry out some Granger causality analysis. One speci c use we make of the signi cance tests is to omit from thevars theoil variables which arefound tobenot statistically signi cant. With regard tothe Grangercausality analysis, we report results obtained in both a bivariate and a multivariate context. In subsection 3.2 we formally test the possible existence of nonlinearity throughout the nonlinearity test proposed by Hamilton (2001). We next turn in subsection 3.3 the examination of the e ects of oil price shocks on GDP growth. We present the results on impulseresponse functions and accumulated responses obtained using all linear and nonlinear speci cations. When discussing the results, we distinguish between net oil importing and exporting countries. After presenting the results, we compare the performance of the di erent speci cations considered here. In subsection 3.4 we focus on the single most important equation of the VAR systems, namely the GDP equation, testing for the stability on oil price coe cients. 8

9 3.1 Testing signi cance and Grangercausality Signi cance Under this item we investigate the signi cance of the impact of real oil prices on real activity and the other variables of the model (in both its linear and nonlinear speci cations) by means of two di erent tests. First, the Wald test tests the null hypothesis that all of the oil price coe cients are jointly zero in the GDP equation of the VAR model. 19 Second, the Likelihood Ratio (LR) test tests the null hypothesis that all of the oil price coe cients are jointly zero in all equations of the system but its own equation. Therefore, the LR test is a test of the signi cance of oil prices for the VAR system as a whole. This means that if oil prices were found not to be signi cant from the LR test, they could be eliminated from the system of equations. The LR test is also informative in that it could well be that oil prices do not a ect GDP directly (as assessed by the Wald test), but through third variables in the system. The results on the Wald test indicate that we cannot reject the hypothesis that the alternative oil prices measures (linear and nonlinear) are not statistically signi cant at a 5% critical level in most of the countries considered. This means that oil prices do not appear to have a signi cant direct impact on real activity. The main exception is the euro area as a whole, in which case all of the nonlinear oil price measures are statistically signi cant at a 5% critical level. 20 The LR test results are indicative of the signi cance of oil prices for the system in most of the countries. In the two speci cations where we do not distinguish between oil price increases and decreases, namely the linear and the net speci cations, the oil price variable is signi cant at the 5% level, with the exception of the US in the linear model. The latter result is consistent with the concerns expressed in the literature about the poor performance of US linear models since the mid1980s. 21 In the other two speci cations, namely the asymmetric and scaled models, the results show that theoil price increase variable is signi cant in all countries, while the oil price decrease variable is not signi cant in most countries. The only cases where the latter variable is signi cant is in Canada and the UK for both models, and in the US for the asymmetric model. 22 The fact that the declines in oil prices are found to be statistically signi cant in some of the countries 19 Some authors, like for instance Hooker (1999) in the related literature, call this test a multivariate Granger causality test. In the next subsection of this paper we use that name in a di erent way (see Hamilton, 1994). 20 The other exceptions in which some oil price measures were statistically signi cant at a 5% critical level are thesopi t andnopit in Germany and thesopdt in Canada. 21 Already in the late 1980s, Mork (1989) who also employed a multivariate framework found that the oil price coe cients were not signi cant at the 5% critical level in his US GNP equation, showing only borderline signi cance at the 10% level. 22 Similar results for Canada and the US are found by Mork et al. (1994) using a somewhat di erent version of the asymmetric speci cation. The nding that oil price decreases are found to be signi cant for the US economy, with the implication that we shall thus explicitly consider the impact of the negative oil price measure, contrasts with the approach of most of the studies in the literature who simply omit consideration of oil price decreases in their models. 9

10 considered suggests that the omission toconsider a priori such movements, as doneby most of the studies in the literature, seems to be a questionable and inappropriate simpli cation. In sum, we found signi cance of oil prices in the linear model for all countries but the US, the positive movements in both the asymmetric and the scaled speci cations, and the NOPI speci cation. We also detected signi cance of negative movements in the asymmetric models for the US, the UK and Canada, and in the scaled speci cation for the UK and Canada Grangercausality results: Bivariate and multivariate tests In a multivariate context, the notion of Granger causality needs take into account not only the direct impact of one variable of the system on another, but also its indirect e ect through third variables. In this regard, the concept of Granger causality is assessed in terms of the socalled tests of block exogeneity. In this section, we investigate primarily two aspects of Granger causality in the di erent speci cations of the VAR model. First, we test for whether the oil price variables are Grangercaused by the remaining variables of the system. The results of this test for each country are reported in the rst lines for each country in Table 2. Second, we test whether oil price variables Grangercause the remaining variables of the system, the results being showed in the second lines for each country in Table Concerning the latter test, while our focus is on its multivariate aspects, we also report in Table 3 for comparison the pvalues of the standard bivariate Granger causality test from real oil price variables to real GDP growth. Finally, in addition the previous tests, we discuss here one additional test, the one for the lack of any relationship between oil prices and the rest of the system, which is reported in the third lines for each country in Table 2. The rst lines for each country in Table 2 allow us to generally reject the null hypothesis that real oil price variables are not Grangercaused by the remaining variables of the system at the 5% signi cance level. The exceptions this are given by the linear models of the US, Canada (where we can reject the null at the 10% level), and Italy, the net model of Canada, and all models in Norway. The second lines for each country in Table 2 show that we can generally reject the null hypothesis that real oil price variables do not Grangercause the remaining variables of the system at the 5% signi cance level. The exceptions this are the linear models of the US and the euro area, as well as the scaled and net speci cations for theus. We now turn tothe description of the standard bivariate Granger causality 23 The price decrease variable was consequently eliminated from those asymmetric and the scaled speci cations in which it was found not to be signi cant. 24 In the cases of the lineal and net models, the latter test tests a similar null hypothesis that of the LR test discussed in the last subsection. In the cases of the asymmetric and scaled models, the new test di ers from the previous LR test in two di erent ways. First, the new test tests for joint signi cance of the positive and negative oil price variables, as opposed the separate tests in the previous subsection. Second, for those countries for which the negative oil price variables were eliminated in line with the LR tests, the new test only tests for signi canceof the positiveoil price measures. 10

11 tests from real oil price variables to real GDP growth, which are reported in Table 3. In the cases of the linear modeland theoilprice increase measures ofthe nonlinear approaches, theresults indicatethat we reject the null hypothesis that oil price measures do not Grangercause GDP growth in the US, the euro area and Italy. Instead, we cannot reject the null in the cases of Germany, Japan, Canada and Norway. In the cases of France and the UK the results di er depending on the model considered. 25 Furthermore, we accept in all models the null hypothesis that oil price decrease variables do not Grangercause GDP growth. Finally, the third lines for each country in Table 2 permit us to reject the hypothesis that there is no relationship between oil prices and the rest of the system in all countries but Norway results which are robust across model speci cations. Consideration of all tests reported in Table 2 suggest that the result for Norway is driven by the already mentioned lack of response of real oil prices the remaining variables of the system. In sum, the results show that the interaction between oil price variables and macroeconomic variables is found to be signi cant, with the direction of causality going in at least one direction in all countries and in both directions in most countries. 3.2 Nonlinearity test Hamilton (2001) develops atest of the null hypothesis that the truerelationship between two variables is linear against abroad class of nonlinear alternatives based on the Lagrange multiplier principle. We have performed such a test to assess the possibility of nonlinear relationship between real GDP growth and real oil price changes in each of the countries under study. On observing Table 4, we nd that we cannot reject the null hypothesis of linearity in all countries. This result, however, should be considered with caution given that JimenezRodriguez (2002a) observed 26 that the null hypothesis of linearity using a subsample of US data that runs from 1972:III to 2001:III cannot be rejected, despite the fact that the underlying relationship between the two variables is nonlinear. As such, it makes sense to consider the nonlinear speci cations for all other countries under study apart from the US. 25 In France we accept the null at the 5% signi cance level (but we reject it at the 10% level) in the case of the linear model, while we reject it for all oil price increase measures. In the UK we accept the null at the 5% level in all cases but the net speci cation (although we reject the null at the 10% level in the case of all oil price increase variables). 26 She also pointed out that the power of such a test could be low when there is a structural change in the marginal distribution of the regressors (see the discussion in JimenezRodriguez, 2002a). 11

12 3.3 Macroeconomic Impacts of Oil Price Shocks: Linear and Nonlinearspeci cations In this subsection, we assess empirically the e ects of oil price shocks on economic activity by analysing the impulseresponse functions and their corresponding accumulated responses in the context of both linear and nonlinear speci cations. We then assess the relative performance of the di erent models Linear speci cation Under this item we examine the e ects of oil prices on GDP growth in terms of both orthogonalised impulse response functions and accumulated responses for the linear speci cation of the model. When describing the results, we distinguish between the results for net oil importing as opposed to exporting countries. Figure 2 represents the orthogonalised impulse response functions of GDP growth to one standard deviation oil price shock with their corresponding two standard error bands. In turn, Table 5 ( rst ve lines for each country) reports the accumulated responses of GDP growth to an oil price shock normalised to correspond to a 1% increase in the linear model. 27 In order to better understand the mechanisms behind the impulse and accumulated responses of GDP growth, we have analysed impulse and accumulated responses of other variables. We nd that one of the key channels playing a role in the e ect of oil prices on real activity is related the real e ective exchange rate. Table 6 reports accumulated responses of REER to an oil price shock normalised to correspond to a 1% increase in the corresponding current oil price measure under study, the Table s rst ve lines referring the case of the linear model. While we do not report any gures or tables concerning the oil shock e ects on other variables, we summarily discuss some of these e ects at the end of the present subsection. Results for net oil importing countries: Despite the fact that each of the countries responds somewhat di erently to an oil price shock, Figure 2 permits us to observe a similar pattern of impulse response function. In fact, we observe that the real impact of oil prices is negative in the shortterm with the only exception of Japan 28 where it is positive. 29 The largest negative shortrun in uence takes 27 In the linear model, the optimal lag length was found to be four for all countries but France, Italy and Norway, where the appropriate lag length is three. 28 Mork (1994) reports the annual real growth rates in the leading world economies, including Japan. Observing his Table 1, we observe that Japan was the least hurt of those countries after the Yom Kippur War in The exceptional behaviour of Japanese output growth after the shock is not robust to changes in the order of the VAR. Indeed, when we use a VAR with two lags, we nd that the largest shortrun response of GDP growth to oil shock is negative and takes place during the third quarter after the shock (see Figure 6). The previous results, considering a 2ndorder VAR, are consistent with those of Lee et al. (2001). These authors, using a monthly VAR and a number of lags from 5 to 7 depending on the measure of oil price used, nd that oil price shocks have a negative impacts on output (condering industrial production as the output variable). It should be noted that these monthly lag lenghts correspond to 12

13 place within the year of the shock, being reached in the fourth quarter after the shock in all countries but France, where the maximum negative e ect occurs during the third quarter. The impact of the shock becomes very small after the rst year, dying out almost completely after three years. Table 5 reports that an oil price shock has a negative accumulated e ect on GDP growth in all oil importing countries but Japan. 30 The largest negative accumulated e ect of an oil price shock is on the US economy. Indeed, the accumulated loss of GDP growth after a 1% oil price shock in the US is 0.032%, at least the double of that observed in most of the other net oil importing countries. One important mechanism that helps to explain the larger negative e ects of an oil price hike on US GDP growth is that the US is the only oil importing country for which the linear model yields a real exchange rate appreciation after such a shock (see Table 6). Euro area countries Germany, France and Italy also exhibit relatively large accumulated real impacts of a positive oil price shock, which is consistent with a considerable oil dependency. The real e ective exchange depreciation induced by the shock o sets only in part the negative impact on output growth predominant in these countries. 31 In turn, Canada exhibits the smallest negative accumulated e ect of an oil price shock, which is consistent with the fact that this economy has been not so dependent on oil before becoming a net exporter of this commodity in the 1980s. Results for net oil exporting countries: Turning the results for net oil exporting countries, Figure 2 indicates that there is a similar general pattern in the reaction of British and Norwegian real GDP growth to oil price shocks within the rst year, with a positive response during the rst two quarters followed by a negative response during the two following quarters. After the rst year, the impulse responses di er somewhat in the two countries. Related this, Table 5 shows that there is a striking di erence between both countries in terms of the accumulated impact after the rst year, with positive overall impacts in Norway and negative impacts in the UK. This means that the UK exhibits an anomalous behaviour: while it is expected that an oil price shock has positive e ects on the GDP growth for a net oil exporting country, an oilprice increase of 1% actually leads toa loss of British GDP growth rate of more than 0.017% after the rst year. An extensive literature has highlighted that this anomalous result has to do with the fact that positive oil shocks led to a large real exchange rate appreciation in the pound (Dutch disease), a fact that is captured in the results presented in Table 6. In connection with this, Table 6 also shows that the Norway s real exchange rate appreciation after the rst year is much a two quartely lags. Nevertheless, the optimal number of quarterly lag for the Japanese economy should not be two, but rather four. 30 As with the impulse responses, the accumulated result is not robust to changes in the order of the VAR. The secondorder VAR yields anegative accumulated impact of oil shocks in JapaneseGDP growth, even though the loss of growth is in this case relatively small (slightly less than after 49 periods). 31 The accumulated impulse response for the euro area as a whole shows a much smaller negative real impact of the oil shock. This may be partlyattributed toeconometric problems relatingtothe aggregation of the transmission mechanisms of the di erent individual countries which formed the euro area only at the end of our sample. 13

14 weaker than in the British case, thereby lessening the positive impact of oil shock on Norwegian GDP growth by a considerably smaller amount. E ects of oil prices on other variables: With regard the e ects of oil price increases on variables other than GDP growth and REER, we nd that an oil price hike induces an accumulated increase in the level of in ation in all countries except Germany, 32 and an accumulated rise in both the short and longterm interest rates in all countries but Canada, Germany and the euro area. 33 With regard the growth rate of real wages, we observe an accumulated decrease in all countries but Norway. 34 These results are plausible and provide evidence of transmission mechanisms other than the exchange rate channel playing the expected roles in most countries. In sum, the linear model indicates, on the one hand, that an oil price increase leads to a decline in GDP growth rate in all oil importing countries but Japan. On the other hand, the linear model provides contrasting results for the two oil exporting countries in our sample as it indicates that an oil price hike has a positive e ect on Norwegian GDP growth, but a negative impact on British GDP growth (related to Dutch diseasetype e ects) Nonlinear speci cations We now turn the results obtained for nonlinear speci cations of the model, again distinguishing between net oil importing and exporting countries. Figures 3.1 through 5 represent the orthogonalised impulseresponse functions of GDP growth to one standard deviation oil price shock, while Table 5 (last fteen lines for each country) reports the accumulated responses by GDP growth to an oil price shock normalised to correspond to a 1% increase in the corresponding current oil price measure under study. 35 The last fteen lines for each country in Table 6 report accumulated responses of REER to an oil price shock normalised to correspond to a 1% increase in the corresponding current oil price measure under study. We describeheretheresults for the threenonlinear speci cations, namely the asymmetric, scaled and net approaches, at the same time. The reason for this is that the results tend to be qualitatively similar. As with the linear speci cation, we distinguish here between the results for net oil importing as opposed toexportingcountries. Moreover, when describing the results we distinguish between oilpricemovements 32 While the level in ation does rise in Germany up the fth quarter, it thereafter decreases in accumulated terms. 33 In these three cases both interest rates actually fall after one year. 34 In the Norwegian case the rate of growth in real wages actually increases. This is alshe case in Canada between the third and the sixth quarters after the shock. 35 In the case of the nonlinear models, the optimal lag lengths remain the same as in the linear speci cation for all countries in which the oil price measures are statistically signi cant. The only exception the latter statement was Canada in the case of the NOP I speci cation, wherethe optimal lagis no longer four but three. 14

15 relating to increases and decreases. The main motivation for this distinction is the previous nding in subsection that the negative movements of oil prices are not statistically signi cant in most of the countries under study. In terms of the results to be described, this has the two following implications. First, with regard the impulseresponse functions, it means that we shall describe rst the e ects of positive oil measures, that is, o, SOPI, and NOP I, for all models and all countries (as captured by Figures 3.1, 4.1 and 5) and those of the negative oil measures for the asymmetric and scaled speci cation in only a few countries (see Figures 3.2 and 4.2). 36 Second, as far as the accumulated responses are concerned, the relevant lines in Tables 5 and 6 contain empty cells in the case of those countries for which the negative oil price measure was previously found to be not signi cant. Results for net oil importing countries: In the case of positive movements in oil prices, the results for net oil importing countries indicate that the results are qualitatively comparable those of the linear speci cation. One exception concerns the exact timing for the largest negative real impact of the oil shock in the case of Italy. In fact, while the largest negative impact on Italian GDP growth occurs in the fourth quarter in the linear case, it takes place in the third quarter in the nonlinear speci cations. In addition to being qualitatively similar the linear speci cation, the nonlinear models yield shapes of the impulseresponses of GDP growth to oil price hikes that are roughly comparable across nonlinear speci cations, except for the timing of largest negative real impact of the oil shock in two cases. The rst exception is given by the asymmetric speci cation in the US, which presents positive movements in GDP growth during the rst three quarter and a maximum negative impact in the fth quarter. This contrasts with negative output developments in the rst year, with a maximum negative impact in the fourth quarter,in the other nonlinear speci cations afeature shared alsobythelinear model. Secondly, in Canada, while the maximum negative real impact is in the fth quarter in net model, the other nonlinear speci cations (as well as the linear model) have the fourth quarter as that of the maximum negative impact on GDP growth. 37 Turningnow totheaccumulated responseofgdp growth toapositiveoilpriceshock, Table5indicates that they are qualitatively comparable those found in the linear case. This includes the nding that the US is the economy that exhibits the most negative e ects on GDP growth, associated with a real exchange rate appreciation in e ective terms. 38 Moreover, as with the linear model, the nonlinear speci cations 36 We have alsested for the null hypothesis that positive and negative coe cients are equal in the VAR framework, obtaining the rejection of null hypothesis in all cases. For this reason, in the cases where the negative movements are signi cant we consider positive and negative oil prices as separate variables. 37 The di erence between the impulse responses in the fth and the fourth quarter under the net speci cation is, however, very small. 38 See Table 6. The fact that the Canadian dollar exhibits a real e ective appreciation represents a di erence with respect the linear approach. Among European net oil importers, the euro area as a whole and France exhibit a real e ective exchange rate depreciation which is robust across nonlinear speci cations (and also observed in the linear model). In Italy 15

16 yield the anomalous result of a positive reaction of Japanese GDP growth to an oil price increase. 39 With regard the magnitude of the accumulated response, the nonlinear speci cations yield larger negative impacts on GDP growth than in the linear case, with the SOP I variant showing a larger impact than the NOP I approach and the latter in turn producing larger e ects than in the o case. 40 Turning now the case of decreases in oil prices, we nd evidence of a signi cant negative impact on US real GDP growth according the asymmetric speci cation a result that contrasts with the earlier literature which found no evidence of any signi cant e ect. The positive impulse responses for GDP growth start after the rst year, contributing turn positive the mediumterm accumulated e ects. However, the signi cance of this e ect vanishes when we look at the results obtained using the scaled model a somewhat better speci cation in terms of the analysis done below. In Canada, the asymmetric and scaled approaches both yield the surprising result that a lower oil price induces a fall in GDP growth and a real e ective exchangerate appreciation. One possible interpretation of this anomalous result is that thelargest oil price decreases occurred after theearly 1980s, that is, after the country switched from net oil importer into net oil exporter. Results for net oil exporting countries: With regard to positive oil changes, the results for net oil exporting countries show asimilar pattern across thedi erent nonlinear speci cations within the rst year, which are alsoqualitativelysimilar the linear speci cation. In particular, we observethat in both theuk and Norway real GDP growth shows apositive response tooilprices during the rst twoquarters followed by anegative response during the two followingquarters. Impulse responses di er between both countries after the rst year, which leads the di erence in Table 5 in terms of the accumulated impact beyond the short term, with overall a positive real impacts in Norway and negative real impact in the UK. In both the British and Norwegian cases, the SOPI speci cation yields accumulated impacts on GDP growth after three years of similar magnitude but with di erent sign, which is for both this model and the N OP I model considerably larger than the overall e ect captured by the asymmetric approach. As we discussed in the case of the linear model, the peculiar result for the UK can be rationalised in terms of the nding that positive oil shocks lead to a large real appreciation in the pound, as presented in Table 6. In connection with this, Table 6 also shows that Norway s real exchange rate appreciation after the rst year is much weaker than in the UK, thereby having a much smaller negative partial e ect on real GDP growth. and Germany, the sign of the accumulated impact on REER depends on the nonlinear speci cation in question. 39 As with the linear model, this result is not robust to changes in the number of lags used. Again, the secondorder VAR yields a negative accumulated impact of oil shocks in Japanese GDP growth. 40 The only exception is that of the net speci cation for Canada, where the losses of GDP growth rate during the rst one and a half year after the shock are smaller than in the linear case. However, it is worth noting that Canada is a country where the negative movements of oil prices are statistically signi cant in the other nonlinear models, indicating that the net speci cation may omit valuable information. 16

17 Regarding decreases in oil prices, which is only signi cant in the case of the UK, Figures 3.2 and 4.2 show that the shapes of the impulseresponse functions of GDP growth to changes in the o and SOPD variables are largely similar. In accumulated terms, Table 5reports that both the asymmetric and scaled models exhibit a positive accumulated e ect, 41 which is somewhat larger in the latter case. Moreover, comparison between the impacts on British GDP growth of oil price increases and decreases show that the latter have a much larger e ect (about three times as large) using both the asymmetric and scaled speci cations. E ects of oil prices on other variables: The accumulated e ects of oil prices on variables other than real GDP growth and REER are broadly similar across nonlinear speci cations. An increase in oil prices induces an accumulated increase in the level of in ation in all countries but Germany, and an accumulated rise in both the shortterm and the longterm interest rate in all countries except Germany. With regard the growth rate of real wages, we observe an accumulated fall in all net oil importing countries but the euro area and Canada (and also Japan in the net speci cation), and an increase in net oil exporting countries with the exception of the UK in the net speci cation. Comparison with results of the linear model: Using the linear approach, we are assuming that there are symmetric e ects of an increase and a decrease in oil prices on real economy. That is, oil price declines are considered tobe as bene cialas oilprice increases of the same magnitude are detrimental to real economic activity. The e ects of oil price increases on GDP growth using the nonlinear approach, on the onehand, tend to be qualitatively (although not quantitatively) similar those of the linear one, reducing GDP growth in all oil importing countries but Japan and oil exporter UK and raising GDP growth in Norway. The e ects of decline in oil prices, on the other hand, di er substantially between the two approaches. In the linear case these e ect are equivalent those of increases, however when we consider the nonlinear approach such e ects are nonconsidered in the net speci cation and are found to signi cant in only few countries in the asymmetric and scaled speci cations. Speci cally, the asymmetric and scaled models allow us to have an independent assessment of the impact of oil price declines, which turns out to be positive in the US and the UK, but negative in Canada. With regard the magnitude of the e ects of oil shocks produced by nonlinear models, it is generally larger than in the linear approach, particularly in the case of the scaled speci cation that takes the volatility of oil prices into account. In order to provide some evidence on the relative performance of the di erent models, we report in Table 7 the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) for all countries obtained from each speci cation. On the basis of these two criteria, we observe that the scaled 41 Again, in light of the real e ective depreciation documented in Table 6, the impact of a fall in oil prices on real GDP growth can be interpreted as a Dutch disease e ect operating in reverse. 17

18 speci cation performs in all countries somewhat better than the other approaches used in the present study Stability Tests In this subsection we investigate the stability on oil price coe cients of the GDP equation for each country in each of the four speci cations used in the paper. To do so, we calculate the Andrews s test (1993) and Andrews and Ploberger s tests (1994), the results being reported in Table 8. We de ne the rst part of the sample to be T 1 = [¼T] and the last part of the sample to be T 2 = [(1 ¼)T]; where ¼ is some arbitrary fraction that we set equal to 30%, which determines that the period between T 1 and T 2 lies between 1981:III and 1993:II. We then look for a possible breakdate t 1 between T 1 and T 2. Let the GDP equation of the VAR be written in the following way: y t = a 0 0 x 0t 01x 1t 02x 1t ± [t>t1 ] 0 x 2t " t (4) where y t is the real GDP growth; x 0t and x 1t are pdimensional vectors which contain lags in y t and in oil price growth, respectively; x 2t is a ((n 2) p) 1 vector which contains lags in all of the other variables of the VAR model. 43 Let F (t 1 ) be the Wald test of equality of the coe cients 1 and 2 under the null hypothesis that there is not a structural break, that is, F(t 1 ) = (T k)(rss 0 RSS 1 ) RSS 1» a  2 p where RSS 0 is the residual sum of squares from OLS estimation of (4) under the null hypothesis of no structural break, and RSS 1 is the corresponding to unconstrained OLS estimation of (4). Andrews (1993) proposes test for the presence of a break point by the following test statistic: supf = sup F (t 1 ) T 1 t 1 T 2 If the observed test statistic passes the critical value, then the date t 1 that satis es sup F(t 1 ) will be T 1 t 1 T 2 theestimated breakdate. Likewise, Andrews and Ploberger (1994) suggest two alternatives test statistics to verify the existence of a break point: XT 2 Avg F = (T 2 T 1 1) 1 t 1=T 1 F(t 1 ) 42 Hamilton (2003) also favours the scaled speci cation in his study of the US economy. JimenezRodriguez (2002b) nds that a kernel semiparametric speci cation could improve even upon the scaled speci cation in the case of the US. As we indicated in the Introduction, the estimation of a kernel model requires a longer sample that the one available for most of the countries represented here, and for this reason this methodology is not considered in the present paper. 43 Notice thatt1 is a nuisance parameter that appears under the alternative hypothesis, but not under the null hypothesis. 18

19 " # XT 2 Exp F = ln (T 2 T 1 1) 1 exp(f(t 1 )=2) t 1 =T 1 The results in Table 8 suggest that, in general, there is no evidence of instability of oil price coe cients at a 5% signi cance level for any speci cation and for all countries. The only exception is the linear speci cation for the euro area, where we nd evidence of instability in 1981:IV. 4 Concluding Remarks This paper studies the e ects of oil price shocks on the real economic activity of the main industrialised OECD countries, distinguishing between net oil importing and exporting countries. Our interest is focused on the relationship between oil prices and real GDP growth, which is analysed in terms of vector autoregressions. We use four speci cations of the latter, namely a linear model and three leading nonlinear speci cations proposed in the literature. Our ndings suggest that the e ects of an increase and a decrease in oil prices on real economy di er substantially, showing evidence against the linear approach that assumes that oil price declines are as bene cial (detrimental) as oil price increases of the same magnitude are detrimental (bene cial) to real economic activity. The results obtained in the paper are broadly consistent with the expectation that the real GDP growth of oil importing economies 44 su ers from positive oil shocks. With regard to net oil exporters, Norway bene ts from oil price hikes while in the UK a rise in oil prices is found to have a signi cant negative impact on GDP growth. This result relates the evidence of adutch disease e ect. Furthermore, the use of twypes of nonlinear models, namely the asymmetric and scaled speci cations, allow us to distinguish between the e ects of oil price increases and decreases. By using these approaches, we nd that a decline in oil prices only a ects signi cantly some countries. For instance, we observe that a fall in oil prices have a positive impact on the US economy after the rst year following the oil shock and on the British economy, while it has a negative e ect in Canada. The latter result can be explained by the fact that large decreases in oil prices took place after the early 1980s, which is when the Canadian economy was switching from a net oil importer into a net oil exporter. With regard the size of the e ects of oil shocks, the largest real e ects of oil price increases were found take place in the case of the US economy. This is particularly the case when we look at the results of the linear model. Nonlinear speci cations indicate that euro area countries (Germany, France 44 In the case of net oil importer Japan, the results obtained using the optimal order of the VAR model (i.e. four lags) indicate a positive association between oil prices and real performance. This anomalous result can be rationalised in terms of the peculiar circumstances undergone by the Japanese economy during the period under study. Indeed, this economy proved rather resilient the oil shocks of the 1970s and early 1980s despite its large dependence on oil, while it failed to bene t from the declines of oil prices observed since the mid1980s. Furthermore, the anomalous result is not robust, being reversed in the case of a secondorder VAR. 19

20 and Italy) seem to be a ected in a magnitude comparable that of the US. Additionally, we found non evidence of instability on oil price coe cients of the GDP equation for each country considered, with the only exception of the linear speci cation in the case of the euro area. 20