The U.S. Dollar Exchange Rate and the Demand for Oil

Using recent advances in panel data estimation techniques, we find that an appreciation of the US dollar exchange rate leads to a significant decline in oil demand for a sample of 65 oil-importing countries. The estimated effect turns out to be much larger than the impact of a shift in the global crude oil price expressed in US dollar. Furthermore, the effect of the US dollar on oil demand tends to be declining over time and, for a subsample of OECD countries, stronger for an appreciation compared to a depreciation of the US dollar.


Introduction
There is a growing consensus that global crude oil price fluctuations are mainly driven by changes in the demand for oil. Hamilton (2009)  Surprisingly, the role of the US dollar exchange rate has so far been ignored in these empirical studies. In particular, since global oil prices are predominantly expressed in US dollars, a shift in the dollar exchange rate should affect the demand for crude oil in countries that do not use the US dollar for local transactions. When the US dollar exchange rate depreciates, oil becomes for instance less expensive for consumers in non-US dollar regions, boosting their demand for oil (Austvik 1987). The rise in oil demand outside the US should in turn influence global oil production and oil prices expressed in US dollar.
In other words, changes in the US dollar exchange rate could be an important source of global oil demand fluctuations. This conjecture is supported by the data shown in Figure   1. The figure shows the evolution of the real effective US dollar exchange rate, together with respectively detrended global oil production (right panel) and the real price of crude oil expressed in US dollars (left panel). As can be seen in the figure, an appreciation (depreciation) of the dollar exchange rate is typically accompanied by a decline (rise) in both global oil prices and oil production, indicating a fall (rise) in global oil demand.
The above mentioned studies, however, do not explicitly consider shifts in the US dollar exchange rate as a possible driver of global oil prices, production and consumption.
A similar argument could be made for several studies that exclusively focus on the analysis of the determinants of oil demand. In particular, Gately and Huntington (2002), Cooper (2003), Dargay, Gately and Huntington (2007), Narayan and Smyth (2007) and Dargay and Gately (2010) amongst others estimate oil demand functions for multiple countries. These studies consider oil demand as a positive function of income per capita and a negative function of its own price. For the latter, they typically use global crude oil 2 prices expressed in US dollars. The influence of shifts in the US dollar exchange rate on oil demand is not taken into account in these studies, which could even bias the estimated income and price elasticities. Some studies (e.g., Griffin and Schulman 2005;Fawcett and Price 2012) do use local oil (gasoline) prices in the estimations, but do not distinguish between local oil price movements that are caused by global oil price shifts or by changes in the exchange rate. There is, however, no a priori reason to assume that the influence of both sources of oil price shifts on oil demand is the same.
In this paper, we formally examine the role of the US dollar exchange rate as a driver of oil demand in non-US dollar regions. Specifically, we estimate the determinants of oil consumption per capita for a panel of 65 oil-importing countries over the sample period 1971-2008. A panel data approach is commonly used in the literature on oil (energy) demand, as it allows to exploit both the cross-section and the time dimension of the data to identify the effects. We conduct panel estimations for respectively a sample of 23 OECD countries, 42 non-OECD countries and all 65 oil-importing countries. Besides real GDP per capita, we include global real crude oil prices expressed in US dollar, as well as the real effective US dollar exchange rate in the estimations. An explicit analysis of the role of the US dollar as a possible driver of oil consumption is a first contribution of the paper. 1 A second contribution of the paper is methodological. In particular, most existing panel data studies on oil demand do not fully take into account the specific salient features of macro panel data sets such as heterogeneity of the coefficients, unit root behavior and crosscountry dependence, even though the neglect of these matters can result in misleading estimation outcomes. We apply recent advances in panel estimation techniques that are capable to handle these econometric issues. Specifically, we (i) take into account the cointegration relationships between the variables by estimating a panel error correction oil demand model, (ii) allow for cross-country heterogeneity of the coefficients which is present in the data, and (iii) correct for cross-sectional dependence in the error terms. The application of these econometric advances turns out to matter for some of the estimated coefficients.
We find that an appreciation of the US dollar real effective exchange rate leads to a decline in oil consumption in non-US dollar regions. Strikingly, the short-run US dollar exchange rate elasticity of oil demand turns out to be much larger than the elasticity of oil demand with respect to fluctuations in the world price of crude oil expressed in US dollar. A back-of-the-envelope calculation furthermore suggests that shifts of the US dollar exchange rate are economically an important contributor to the volatility of the global price of crude oil expressed in US dollar, due to its influence on oil demand. These findings underline that the US dollar exchange rate should be taken into account in the analysis of global oil market dynamics and sources of oil price fluctuations.
This paper also highlights the existence of time variation and asymmetries in the reaction of oil demand to US dollar exchange rate changes, which can be considered as a third contribution. On the one hand, in line with Hughes, Knittel and Sperling (2008) and Baumeister and Peersman (2012), who show that the price elasticity of oil demand has declined considerably since the mid-eighties, we find a similar time shift in the response of oil demand to the dollar exchange rate. On the other hand, we find that oil consumption in OECD countries reacts asymmetrically to a strengthening or weakening US dollar, with a stronger reaction to an appreciation of the US dollar. This finding is consistent with Gately and Huntington (2002) and Huntington (2010), who document an asymmetric impact of rising versus falling oil prices on the demand for oil.
The remainder of this paper is organized as follows. In the next section, we describe the baseline empirical model for oil demand and discuss some econometric issues. Section 3 discusses the estimation and robustness of the results, while the economic relevance is assessed in section 4. Nonlinearities of the exchange rate effects are examined in section 5. Finally, section 6 concludes.

Empirical oil demand model
In this section, we describe the benchmark oil demand model that will be used in the estimations. Our sample contains 65 oil-importing countries and covers the period 1971-2008. 2 Details of the data and a list of the countries can be found in Appendix A. Consider 2 We do not consider oil-exporting countries since oil demand in these countries have been found to behave very differently. See for example Gately and Huntington (2005). 4 the following general oil demand specification for country i at time t: where dem it is total oil consumption per capita, gdp it real income per capita, oilp t the global real US dollar crude oil price and rer t the real effective US dollar exchange rate, trend t a linear time trend and c i a country-specific constant. All variables are converted to natural logarithms, such that the model is of the constant elasticity form. The data are at annual frequency.
The existing empirical literature typically considers oil consumption, or energy consumption more generally, as a positive function of real income and a negative function of its own price (e.g., Dahl (Austvik 1987).
Accordingly, we adopt the global real crude oil price expressed in US dollar, as well as the real US dollar effective exchange rate as two separate variables in our empirical oil demand model. 3 We use the US dollar real effective exchange rate rather than real bilateral exchange rates in the benchmark estimations for three reasons. First, a multilateral weighted exchange rate reflects changes in the value of the US dollar itself, which is the variable of 3 The choice to use global crude oil prices is in the existing literature often motivated by the lack of availability of individual country end-use oil price data, which would constrain the sample considerably (e.g., Gately and Huntington 2002). The use of global oil prices also avoids endogeneity problems of using local oil prices. In particular, the global oil price can be considered as exogenous for individual country's oil demand. Some studies (e.g. Griffin and Schulman 2005) ignore these problems and do use local oil (gasoline) prices for a limited set of countries.

5
interest for the analysis of global oil market dynamics. Second, bilateral exchange rates (or domestic CPI) are not available for several countries over the whole sample period, which would significantly reduce the size of the dataset. Third, bilateral exchange rates suffer an endogeneity problem as the demand for oil of an individual country is expected to influence its own exchange rate. Given the limited weight of each individual country in the US trade basket, this is much less the case for the US dollar effective exchange rate.
In section 3.2, we assess the robustness of the results for a specification with the bilateral exchange rates estimated for a subsample of countries using instrumental variables.
Microeconomic theory (e.g. Mas-Colell, Whinston and Green 2007) suggests that oil demand is also a function of the economy's structure, technology and the prices of substitutes. To capture the former two, all our estimations have a country-specific constant and a linear trend. Unfortunately, prices of substitutes are not available for our sample period. Griffin and Schulman (2005), however, report only weak substitution effects when they include the real price of substitute fuels in a demand model for petroleum products, and argue that omitting cross-price effects does not appreciably alter their results. 4 Notice also that such effects are implicitly taken into account in our analysis since we correct for cross-sectional dependence in the estimations, as discussed in section 2.3. Specifically, when the interest rate declines, commodities become more attractive as an asset for investors.

Panel unit root and cointegration tests
In addition, a lower interest rate stimulates overall demand, including the demand for oil. Notice, however, that this is not relevant for our analysis since we consider oil consumption (not inventories) at the LHS of the oil demand function, while real GDP is included at the RHS. 5 Other panel unit root tests that also use a common factor representation of the data to allow for cross- is determined by the Bai and Ng (2002) criteria. 6 7 Notice that the GUW test is however more restrictive than residual-based tests by imposing weak exogeneity on the country-specific regressors of the ECM and strong exogeneity on the common factors, whereas residual-based tests allow for full endogeneity. To take this restriction into account, we have also applied residual-based panel cointegration tests in the spirit of Banerjee and Carrion-i-Silvestre (2006) to check the robustness of the results. 8 The results of the tests are shown in Table 2. The pooled GUW tests reject the null hypothesis of no error correction between dem it , gdp it , oilp t and rer t for the two models under 6 We consider the IC1, IC2 and BIC3 criteria. The BIC3 criterion is more robust when there is cross correlation in the idiosyncratic errors (Bai and Ng 2002). 7 The common factor critique applies to residual-based panel cointegration tests as they rely on residual rather than structural dynamics (Gengenbach et al. 2008). 8 The approach we take to examine the robustness of the GUW test results is the following: we apply the continuously-updated and bias-corrected (CupBC) estimator of Bai et al. (2009) to the long-run cointegration equation and we test the (de-factored) residuals for a unit root using the PANIC test procedure.
consideration. The alternative residual-based cointegration tests confirm this result. As a consequence, we can safely conclude that dem it , gdp it , oilp t and rer t are cointegrated at the panel level.

Panel error correction oil demand model
Having established cointegration between the variables, we can formulate our general oil demand specification as a panel ECM: Equation (2) is the baseline empirical specification for the panel ECMs that will be estimated in this paper. 9 Gately and Huntington (2002) and Griffin and Schulman (2005) are most closely related to our study as they both estimate single equation total oil demand models for a panel of multiple countries with a moderate time dimension. 10 Before we estimate the panel ECM, it is important to discuss two econometric issues which are typically disregarded in the existing oil demand literature, namely slope heterogeneity and cross-sectional error dependence.

Econometric issues
In order to obtain reliable estimates, we need to consider two important econometric features of macro panel data. The first one concerns heterogeneity in the slope coefficients.
Heterogeneity in the slope coefficients renders the standard FE estimator biased as the latter assumes homogeneity in the slope coefficients. Gately and Huntington (2002) and 9 The lag order of the dynamic adjustment process is imposed to be 0 for all variables for reasons of parsimony. Experiments with more lags, however, do not alter the results. 10 Gately and Huntington (2002)  FE estimator suffers small T problems in dynamic panels (Arellano and Bond 1991). Given that the time dimension of our sample is moderately large, we assume that this bias is not relevant for our purposes.
Nickell (1981) namely shows that the upward bias on the error correction term becomes insignificant when T → ∞. Similarly, Griffin and Schulman (2005), who only analyze a sample of OECD countries, do not assess the homogeneity assumption across OECD countries. Given the cross-country differences in economic structures, the assumption of homogenous slope coefficients is questionable. The homogeneity assumption could, however, easily be tested by means of a Hausman or likelihood-ratio test. Indeed, the application of both tests consistently reveals that the homogeneity restriction on the slope coefficients is not valid, even not for a subsample of OECD countries (see Table 3). The FE estimator is hence biased and the estimated coefficients potentially misleading. Accordingly, we use the Mean Group (MG) estimator in the analysis, which offers a consistent alternative as the MG estimator does not impose homogeneity.
The second important feature of macro panel data estimations relates to error crosssection dependence. In particular, the results of standard FE and MG estimators are inefficient and have biased standard errors when the observed explanatory variables are correlated with unobserved common factors (Pesaran 2006). For oil demand, this is likely the case. Country-specific income, the real price of crude oil, as well as the US dollar real effective exchange rate could for instance be driven by a common global business cycle.
The existing empirical oil demand studies do nevertheless not correct for the potentially far-reaching consequences of cross-sectional dependence. The presence of cross-section dependence in the error terms of dynamic models could be tested with the CD test of Pesaran (2004). Applying the CD test to our panel error correction oil demand model shows that there is a significant degree of cross-sectional correlation in the error terms for both the FE and MG estimators (see Table 3), which confirms our concern about possibly biased results of existing panel studies that do not take into account the dependence.
The common correlated effects (CCE) estimators of Pesaran (2006), notably the CCE pooled (CCEP) and mean group (CCEMG) estimators, are typically used to eliminate cross-sectional dependence. In particular, the CCE estimators use the cross-section averages of the dependent and explanatory variables as observable proxies for a linear combination of unobserved common factors. This approach asymptotically eliminates strong and weak forms of cross-section dependence and is consistent for both stationary and nonstationary common factors (Kapetanios, Pesaran and Yamagata 2010). A non-standard feature of our specification, however, is the presence of observed common factors as explanatory variables, namely the global real price of oil (oilp t ) and the real effective US dollar exchange rate (rer t ), which makes the standard application of the CCE approach of Pesaran (2006) unattractive. 11 We therefore opted to apply the Bai and Ng (2004) (Table 3, bottom row).
In sum, in contrast to the existing empirical evidence on the demand for oil, we do not only examine the role of the US dollar exchange rate for the demand for oil, we also apply panel estimators that take both heterogeneity of the coefficients and cross-sectional dependence into account.
3 Empirical results Table 3 summarizes the estimation results of the panel error correction model as described in section 2. In order to compare with the existing evidence, we report the results for respectively OECD countries, non-OECD countries and the total sample of oil-importing countries. 13 For each sample, we show the results for the FE, MG and MG estimator 11 Adding only the cross-section averages of the real income per capita variables (gdp t−1 and ∆gdp t ) as additional regressors to the model does result in a lower degree of cross-sectional dependence, but the reduction is only marginal. 12 The drawback of this approach, however, is that the estimation error from the first step carries over to the subsequent steps. 13 Given the panel data nature and similar sample of countries, we consider Griffin and Schulman (2005) and the results of Gately and Huntington (2002) without the incorporation of the asymmetric price reactions as the benchmarks.

Panel estimations
adjusted for cross-sectional dependence (MG_Ft), which should allow us to evaluate the relevance of the econometric features discussed in section 2.3 for the estimation results.
Income elasticity We consistently find a significant positive effect of real GDP on the demand for oil. The short-run income elasticity in OECD countries (0.568) is larger than the 0.401 found by Griffin and Schulman (2005), and definitely larger than the zero income elasticity of Gately and Huntington (2002) The econometric issues that we discussed in section 2.3 seem to matter for the magnitudes of the estimates. Specifically, the short-run income elasticity for OECD countries rises from 0.568 to 0.672 if we do not take into account cross-sectional dependence in the error terms, and even to 0.741 if we also do not allow for cross-country heterogeneity in the coefficients. Interestingly, exactly the opposite happens for non-OECD countries, i.e., the short-run income elasticity declines from 0.639 to respectively 0.615 and 0.537 when there is not allowed for correlation and heterogeneity across countries. In other words, the bias resulting from the use of a FE estimator is relevant and works in both directions.
Moving to the long-run income elasticity, there is one caveat that needs to be discussed first. In particular, for some countries in the panel, there seems to be no or only very weak cointegration present in the data. This implies that the estimated error-correction terms for a few countries are very close to zero, which in turn inflates the long-run coefficients for these countries considerably because there has to be divided by the near-zero error correction term to calculate the long-run coefficients. As a consequence, the values of the mean group and the standard errors of the long-run panel coefficients are distorted.
To take this feature into account, we also report the median of the estimated long-run parameters in the tables and interpret them with caution. 14 The median long-run income elasticity turns out to be 0.674, 0.885 and 0.775 for respectively OECD, non-OECD and all 65 oil-importing countries, which is within the range of existing studies.
14 Note that a near-zero error correction term does not influence the short-run elasticities.
(Global) oil price elasticity There are numerous papers that estimate the effects of a shift in (global) crude oil prices on the demand for oil. Most studies report a relatively low, or even an insignificant (e.g. Askari and Krichene 2010) short-run price elasticity of oil demand, which is important because a low oil price elasticity implies that any disruption in oil supply has a considerable impact on the price of oil. As can be seen in Table 3 global oil prices and production. In the next subsection, we assess the robustness of this novel finding.

Some robustness checks
In this section, we first evaluate the robustness of the results for possible endogeneity problems between oil demand, the US dollar exchange rate and the global price of crude oil. It is a standard assumption in existing multi-country oil demand studies that the demand for oil of an individual country does not influence the global price of crude oil on 16 Table 4.
The income elasticity coefficients are comparable to the benchmark results reported in section 3.1. Also the estimated oil price elasticities are only minorly affected. More importantly in the context of the present study, there is still a significant effect of the US dollar exchange rate on the demand for oil. The estimated elasticities turn out to be even larger than the benchmark results, pointing to a possible downward bias of the benchmark results due to the endogeneity of the US dollar exchange rate. The (short-run) exchange rate elasticity is now also statistically significant for non-OECD countries. We obviously have to be careful when interpreting the magnitudes of the coefficients given the loss of power of two-stage regressions with instrumental variables. 17 Nevertheless, these results confirm the significance of the US dollar exchange rate for the consumption of crude oil in non-US regions.
As another robustness check, we have also estimated a specification with bilateral real exchange rates instead of the US real effective exchange rate. While the effective US dollar exchange rate reflects changes in the value of the US dollar, i.e. the currency unit which matters for the global oil market, bilateral exchange rates capture more the effects on the local oil price that consumers have to pay in each individual country. We again opt for IV estimations, with the same instruments as described above. 18 The right part of Table 4 shows the results. Notice that the sample size for the different groups of countries is now smaller due to the non-availibility of bilateral exchange rate and/or domestic CPI data for some countries over the time period under consideration (see Appendix A). This change of sample size should thus be taken into account when comparing the results, especially for non-OECD countries.
The short-run coefficients are comparable to the benchmark results for the OECD sample, in particular the income and (bilateral) exchange rate elasticities. The oil price elasticity, however, turns out to be not significantly different from zero anymore in OECD countries. On the other hand, for non-OECD countries, the income and price elasticity estimates appear to be lower in magnitude compared to the benchmark estimates, whereas the exchange rate elasticity is now larger. Overall, we can conclude from the robustness analysis that the US dollar exchange rate continues to matter for the consumption of crude oil in non-US regions.

Global elasticities and economic relevance
In this section, we evaluate the relevance of our findings for global oil market dynamics.
More precisely, we first calculate global elasticities, which could serve as a benchmark for empirical studies that estimate global oil demand functions based on aggregate oil market data. In a next step, we perform a back-of-the-envelope calculation in order to assess how important fluctuations in the US dollar exchange rate could be for global crude oil production and prices. 18 Note that a possible endogeneity problem between the bilateral US dollar exchange rate and the country-specific demand for oil is much more serious compared to the estimations with the effective ex-

US dollar exchange rate and nonlinearities of oil demand
Having established a significant impact of shifts in the US dollar exchange rate on oil demand, in this section, we examine whether the effects are (non)linear. Specifically, previous work has found a fall in the short-run price elasticity of oil demand over time On the other hand, Baumeister and Peersman (2012) argue that also structural changes in the oil market itself could be a reason for the declined price elasticity of oil demand over time. Baumeister and Peersman (2012) argue that high capacity utilization rates in crude oil production since the second half of the eighties increases the willingness of oil consumers to pay an insurance premium for potential oil scarcity, resulting in a higher share of less elastic speculative or precautionary buying in the oil market. If this structural change in the oil market is the dominant reason for the observed time variation of the price elasticity of oil demand, there is no reason to expect that the demand for oil is subject to a fall in its reaction to US dollar exchange rate changes over time.
In order to examine whether the US dollar exchange rate elasticity has also decreased over time, we add the US dollar real effective exchange rate interacted with a simple linear time trend as an additional variable to the baseline panel ECM: The results of these estimations are shown in Table 6. The first three columns of the

Appreciations versus depreciations
Another popular asymmetry in the price elasticity of oil demand documented in the literature is a stronger effect of oil price increases relative to a decline in oil prices. The basic idea is that higher prices induce more investment in energy-efficient equipment and retrofitting of existing capital, such as switching to energy-efficient cars and greater insulation. When prices fall, however, there is no switch back to less-efficient capital, although there could be more intensive usage, e.g., driving more miles and adjusting thermostats to more comfortable levels (Griffin and Schulman 2005). Gately and Huntington (2002) allow for a different reaction of oil demand to price increases that result in a new historical maximum price, to price increases back to the previous maximum and to price decreases.
They find that oil demand responds differently to the different price shifts, with the largest (negative) effect of price increases that result in new maximum values.
We examine whether there exists a similar asymmetry in the reaction of oil demand to a US dollar appreciation versus a depreciation by adding to the benchmark panel ECM the interaction between the real effective exchange rate with a dummy variable that equals 1 if there is an appreciation of the US dollar effective exchange rate and 0 otherwise: 21 The results are shown in the middle part of Table 6. The interaction term is only significant for the sample of OECD countries. The existence of an asymmetric reaction to a strenghtening versus a weakening US dollar is thus confined to this group. The estimated degree of asymmetry is notably large, i.e., a depreciating US dollar does not invoke a significant oil demand reaction in OECD countries, whereas an appreciating US dollar leads to a considerable decline in oil products demand. The results further reveal that there is no (significant) asymmetry present in non-OECD countries and the total sample of countries. The estimated coefficients have even the opposite (expected) sign. Griffin and Schulman (2005) argue that the results of Gately and Huntington (2002) primarily reflect energy-saving technology rather than price asymmetry. More specifically, they show that the evidence for oil price asymmetry vanishes once time effects that serve as proxies for technological change are incorporated in the model of Gately and Huntington (2002). To take this into account, we have also verified whether the existence of an asymmetric response of oil demand to an appreciating versus depreciating US dollar exchange rate still holds once we control for time variation in the exchange rate elasticity. 22 In particular, we have re-estimated the panel ECM allowing for both forms of nonlinearities.
As such, we could also check whether the decline of the exchange rate elasticity over time, which we have documented in section 5.1, still holds after controlling for an appreciating versus depreciating US dollar. Given that the US dollar experienced a persistent depreciation after its peak level at the beginning of 1985, it is not unlikely that both nonlinearities interfere. These results can be found in the last three columns of Table 6. Again, the interaction term of the dummy variable and the real effective exchange rate regressor is only significant for the OECD sample. Also the trend interaction term coefficients remain 21 Given the validity of the PPP-hypothesis for exchange rates in the long run, we do not distinguish between appreciations of the exchange rate to a new historical maximum and an increase back to the previous maximum. 22 Note that we take the hypothesis of Griffin and Schulman (2005)

Conclusions
In this paper, we have examined the role of the US dollar exchange rate for the demand for oil in non-US dollar regions by using recent advances in panel data estimation techniques. In