Oil Prices and Consumption across Countries and U.S. States

K.7 Oil Prices and Consumption Across Countries and U.S. States De Michelis, Andrea, Thiago R. T. Ferreira, and Matteo Iacoviello Please cite paper as: De Michelis, Andrea, Thiago R. T Ferreira, and Matteo Iacoviello (2019). Oil Prices and Consumption Across Countries and U.S. States. International Finance Discussion Papers 1263. https://doi.org/10.17016/IFDP.2019.1263 International Finance Discussion Papers Board of Governors of the Federal Reserve System Number 1263 November 2019

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1263 November 2019 Oil Prices and Consumption Across Countries and U.S. States Andrea De Michelis, Thiago R. T. Ferreira, and Matteo Iacoviello NOTE: International Finance Discussion Papers (IFDPs) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Oil Prices and Consumption across Countries and U.S. States ∗ Andrea De Michelis† Thiago Ferreira‡ Matteo Iacoviello§ November 4, 2019 Abstract We study the effects of oil prices on consumption across countries and U.S. states, by exploiting the time-seriesandcross-sectionalvariationinoildependencyoftheseeconomies. Webuildtwolargedatasets: onewith55countriesovertheyears1975–2018,andanotherwithallU.S.statesovertheperiod1989–2018. We then show that oil price declines generate positive effects on consumption in oil-importing economies, while depressing consumption in oil-exporting economies. We also document that oil price increases do more harm than the good afforded by oil price decreases both in the world and U.S. aggregates. JEL CODES: Q43, E32, F40 Keywords: Oil prices, consumption, cross-country, U.S. states, oil dependency ∗Erin Markievitz, Lucas Husted, Andrew Kane, Joshua Herman, Patrick Molligo, Brynne Godfrey, and Charlotte Singer provided outstanding research assistance. We are grateful for the comments by Carola Binder. Replication codes for the paper are available here or here. †Federal Reserve Board of Governors. E-mail: andrea.de.michelis@frb.gov ‡Federal Reserve Board of Governors. E-mail: thiago.r.teixeiraferreira@frb.gov §Federal Reserve Board of Governors. E-mail: matteo.iacoviello@frb.gov

1 Introduction The large fluctuations of oil prices in recent years have reignited interest regarding the macroeconomic effects of oil-price shocks. The conventional wisdom of academics, policymakers and market practitioners is that declines in oil prices boost global economic activity, as increases in consumption, especially for oil importers, outweigh the negative effects on oil producers. This wisdom has recently been reiterated by Arezki and Blanchard (2014), Bernanke (2016), and even Warren Buffett, amongst others.1 However, much of the research on the global effects of oil price changes has only looked at a smaller set of countries (see e.g. Jim´enez-Rodr´ıguez and S´anchez, 2005), or has focused on a set of macroeconomic variables, such as industrial production or current account, whose responses may not easily be mapped into those of GDP and consumption (see for instance Aastveit, Bjrnland, and Thorsrud, 2015 and Kilian, Rebucci, and Spatafora, 2009). This paper studies the effects of oil prices on consumption and GDP across countries and across U.S. states. A hurdle in obtaining precise estimates of the global effects of oil-price shocks is data availability. To addressthisshortcoming, weputtogetheraquarterlydatasetcontainingdataonGDP,consumption, andoil dependence for 55 countries. The coverage, which varies across countries, spans more than forty years from 1975 to 2018. We go great lengths to ensure that the data are consistent across countries and do not contain breaks. Using this unique database, we estimate the effects of oil-price shocks on different countries allowing for the effects to vary based on each country’s oil dependence. We find that oil-price declines have large, positive effects on the consumption and GDP of oil-importing countries, while depressing consumption and GDP of oil exporters, and that, in aggregate, the positive effects dominate. However, we also show that the boost to oil importers occurs rather gradually, while the hit to oil exporters is realized fairly quickly, thus suggesting that, in aggregate, the benefits from lower oil prices might occur only slowly. Moreover, we show that oil-price decreases generate smaller boosts to the world consumption than the drag from comparable oil price increases. Weconcludebypresentingcomplementaryanalysisontheheterogeneouseffectsofoilshocksusingstatelevel data for the United States To this end, we use state-level data on registrations of new cars—available quarterlyfrom1989to2018—asaproxyforstate-levelconsumption. Thisdatasethasimportantadvantages over state-level data on personal consumption expenditures, as the latter are still experimental, are available only annually, and cover a shorter time frame. We then show that the effects of oil price declines on consumption differ across states depending on their oil dependence, despite the fact that these states face common monetary and fiscal policy. The paper proceeds as follows. Section 2 presents the results from an aggregate vector autoregressive (VAR) model which groups countries accordingly to whether they are oil importers or oil exporters. Section 3 uses a cross-country panel VAR to dig deeper into the heterogeneous country responses. Section 4 presents the evidence from a state-level panel VAR for the United States. Section 5 relates the results from this paper to the rest of the existing literature on the macroeconomic 1See CBS/AP (29 Feb. 2016). 2

effects of oil shocks. Section 6 concludes. 2 Oil Prices, World Consumption, and GDP Changes in oil prices may reflect disturbances to global aggregate demand as well as disturbances that are specifictotheoilmarket. Inmostofwhatfollows, weinterpretoil-pricechangesasarisingfrom“oilshocks,” controlling for shifts in global economic activity that simultaneously drive oil demand and oil prices. We interpret these oil shocks as reflecting disruptions in oil supply due to geopolitical or natural events, or from precautionary or speculative shifts in the demand for oil. In order to set the stage for the empirical analysis, it is useful to briefly review the key channels by which an “exogenous” decline in oil prices should affect economic activity in both oil-producing and oilimporting countries. On the supply side, lower oil prices raise output in the non–oil sector by reducing firms’ production costs and causing investment and output to rise. This cost channel should be stronger in countries or sectors that heavily rely on oil as an input in production. Conversely, falling oil prices may depress energy-related investment across oil-producing countries, dragging down aggregate activity. On the demand side, lower oil prices transfer wealth towards oil-importing countries and away from oil exporters (e.g, Bodenstein, Erceg, and Guerrieri, 2011), cause a windfall income gain for consumers, and thus shift consumption towards oil-importing countries. This wealth effect may in turn cause GDP to rise through multiplier effects, and may be larger in sectors that produce goods that are complementary to consumption of oil, such as the automobile sector.2 Most of our analysis focus on the consumption effects of oil price changes, and exploit the differential exposure of oil producers and oil importers to these changes to highlight its empirical relevance. Figure 1 shows our first approach to evaluating the differential effects of oil price on consumption accordingtothecountry’sstatusasoilimporterorexporter. Itplotsrealoilpricesgrowthagainstthe“consumption differential,” the difference in consumption growth between importing and exporting countries. Since 1975, wheneveroilpricegrowthhasfallen, theconsumptiondifferentialhastendedtorise, thatisconsumptionhas grownrelativelyfasterinimportingcountries. Thecorrelationbetweenoilpricegrowthandtheconsumption differential is −0.32 and is significantly different from zero. It is important to highlight the consumption differential because it offers a simple way to control for global demand–side effects that otherwise create a positive correlation between activity and oil prices. Indeed, separately taken, the correlations of oil prices with importers’ and exporters’ consumption are both positive, respectively at 0.06 and 0.34. 2.1 Aggregate VAR: Oil Importers vs Exporters Tofurtherexplorethedifferentialresponsesofoilimportersandexporters,weconsideravectorautoregressive (VAR) model aimed at quantifying the effects of shocks to oil prices on global economic activity. The 2See Hamilton (2008) for a discussion on several channels through which oil shocks may affect macroeconomic variables. 3

VAR uses two lags and quarterly data from 1975Q1 to 2018Q4 on oil prices in real terms (oil ), global oil t production (oprod ), importers’ and exporters’ private consumption expenditure (ci and ce , respectively), t t t and importers’ and exporters’ GDP (gdpi and gdpe , respectively). Consumption and GDP are constructed t t aggregating data—using constant-dollar GDP weights—for 45 oil-importing and 10 oil-exporting countries.3 The countries in our sample, as well as their status as oil importer or exporter, are listed in Table A.1.4 The structural VAR representation takes the form: Ax = B(L)x +ε , (1) t t−1 t x = [gdpi ,gdpe ,ci ,ce ,oil ,oprod ](cid:48), (2) t t t t t t t E(cid:0) ε ε (cid:48)(cid:1) = I, (3) t t where B(L) is a lag polynomial of order 2, and E is the expectational operator. We then seek to isolate the effects of oil price fluctuations stemming from oil-market-specific shocks. To doso, weuseaCholeskydecompositionofthevariance-covariancematrixofthereduced-formVARresiduals whereby oil-price shocks affect oil prices and production on impact, and consumption and GDP with a one quarter delay. That is, the reduced form errors can be decomposed as follows:      egdpi a 0 0 0 0 0 ε1 t 11 t    eg t dpe       a 21 a 22 0 0 0 0       ε2 t     eci   a a a 0 0 0  ε3  e t = A−1ε t =   t   =   31 32 33     t  , (4)  ece   a a a a 0 0  ε4   t   41 42 43 44  t        eoil   a a a a a 0  εoil   t   51 52 53 54 55  t  eprod a a a a a a ε6 t 61 62 63 64 65 66 where oil-price shocks εoil impact the economy according to the 5th column of A−1 matrix. This assumption t is in line with a host of empirical VAR studies that place asset prices after the macroeconomic variables in a Cholesky ordering of the residuals of a VAR, on grounds that asset prices are highly sensitive to contemporaneouseconomicnewsorshocks(seeforinstanceStockandWatson(2005),andBernanke,Boivin, and Eliasz (2005)). Figure 2 shows the impulse response functions (IRFs) to an identified oil-price shock that is normalized to imply a 10 percent decline in real oil prices upon the shock impact. This fall in oil prices boosts importing countries’ GDP, with a peak effect reached ten quarters after the shock. Importers’ consumption also rises slowly, with a magnitude slightly larger than GDP. The larger increase in consumption relative to GDP suggests that wealth effects may be an important channel by which oil price declines are transmitted to activity in oil importing countries. By contrast, exporters’ consumption drops rapidly and exporters’ GDP declinesevenmoreso,bottomingoutfourandthreequartersaftertheoil-priceshock,respectively. Thelarger 3The sample includes countries which together contribute to about 80% of world GDP. 4We take the log of all variables, and then quadratically detrend them. See Appendix A.1 for details on data construction. 4

decline in GDP relative to consumption in oil exporting countries is consistent with a rapid deterioration in investment in the energy sector. Finally, oil production moves little on impact and rises persistently in the subsequent quarters, peaking at 0.3 percent above baseline after about two years. Such response supports the interpretation of the oil-price shock as being caused by news about higher future supply. Alltold,theevidencefromtheaggregateVARisconsistentwiththeroleofwealtheffectsonconsumption for oil importers, and suggestive of the importance of direct effects on energy-related investment for oil exporters. 2.2 Aggregate VAR under Different Identification Strategies Our benchmark VAR (equations 1-4) orders oil prices before oil production in the Cholesky decomposition of the variance-covariance matrix of the residuals, with this ordering being often referred as “exogenous oil price assumption” (e.g., Stock and Watson, 2016). This assumption regards unexpected changes in oil prices, once world factors are controlled for, as reflecting international developments specific to oil markets, such as exogenous changes in oil supply conditions. A different approach is to order oil production before oil prices, as in Kilian (2009). With this different ordering, it is possible to distinguish oil-supply shocks—the residuals of the oil production equation in the VAR—from oil-demand specific shocks—the residuals of the oil price equation in the VAR. In this second ordering, oil-demand specific shocks are a key driver of oil prices and are interpreted as shifts in demand due to concerns about future availability of oil supply. Figure 3 shows that the two ordering assumptions identify very similar shocks. In particular, the correlationbetweenouridentifiedoil-priceshocksandKilian’soil-demandspecificshocksis0.41, andisstatistically different from zero. An important reason for the similarity between our estimated shocks and those in Kilian (2009) is that there is little contemporaneous correlation in the data between oil prices and oil production. Therefore, once global conditions are controlled for, it does not seem to matter whether one orders prices before or after production.5 Our VAR ordering does not impose any restriction on the contemporaneous co-movement of oil production and oil prices in response to an oil-price shock. If one interprets the identified shock as stemming from oil supply, the ratio between the IRF at impact of oil production and the IRF at impact of oil prices measures the implied short-run elasticity of oil demand to oil prices. Given the IRFs from our VAR shown in Figure 2, production rises by 0.06 percent when prices drop by 10 percent, thus implying a short-run oil demand elasticity of 0.006, which is in the low end of the estimates of Caldara, Cavallo, and Iacoviello (2019) (henceforth abbreviated to CCI). We then examine the robustness of our benchmark results by adopting an alternative identification strategy. We retain the assumption that oil shocks affect GDP and consumption with a one-period delay, but we decompose the variance-covariance matrix of the VAR residuals assuming joint oil demand and oil supply elasticities (the 2×2 block in the bottom right of matrix A−1) in line with existing studies, such as 5For a detailed discussion, see the appendix in Caldara, Cavallo, and Iacoviello (2019). 5

CCI. In particular, we assume an impact oil supply elasticity of 0.1, which in turn implies a VAR-consistent oil demand elasticity of −0.21.6 As shown in Figure 4, when this newly identified oil “supply” shock is scaled so that its impact on oil prices is the same as that of our benchmark specification, the response of oil production is now larger. Nevertheless, the responses of GDP and consumption of oil importers and exporters are similar to those estimated using the benchmark aggregate VAR. Since the identification strategies from Kilian (2009) and CCI yield results similar to ours, we interpret our results as relatively robust and we maintain a similar order of variables in the VARs of the next sections. Importantly, we focus on identifying oil price movements that are driven by shocks that are specific to the oil market, and thus we do not try to separate out fluctuations driven by demand or supply shocks. 3 Oil Dependency Across Countries: Panel VAR In the previous section, we assumed that the effects of an oil-price shock only depend on whether a country is an oil importer or exporter, disregarding any cross-sectional and time series variations in countries’ oil dependency. Inthissection, weaccountforsuchvariationandshowthatGDPandconsumptionofcountries with higher oil-dependency increase more after exogenous oil price declines. We also show that oil-price increases do more harm than the good afforded by oil-price decreases. 3.1 Oil Dependency Across Countries Figure 5 documents important differences in oil dependence, both across countries, and within country over time. Such differences are obviously lost when we use the simple distinction between oil importers and oil exporters. For a country i at quarter t, we measure oil dependence as the ratio of net oil imports to total oil consumption expressed in percent: ocons −oprod i,t i,t d ≡ 100× , (5) i,t ocons i,t where ocons is the amount of oil consumed and oprod is the amount of oil produced. i,t i,t The top panel (Figure 5a) shows considerable differences in oil dependence across countries in, say, 2013 (similardifferencesshowupwhenlookingatotherpointsintime). Amongoilimporters,Japanandthelarge majority countries in the euro area had an oil dependence of nearly 100 percent, while the United States had an oil dependence of about 50 percent. Among oil exporters, there was even greater heterogeneity according to this metric, ranging from Canada’s (net) oil exports of about 70 percent of its oil consumption to Norway’s 650 percent. The bottom panel (Figure 5b) presents evidence on the variation of oil dependency over time for a select group of countries. For instance, Canada was a net importer in the late 1970s and early 1980s, but now 6Notice that the Kilian (2009) ordering of variables implies a short-run oil supply elasticity of zero, which is not consistent with studies such as CCI. 6

exports a large share of its oil production. Additionally, oil dependency in the United States has fallen sharply since 2010, following the massive growth in shale oil production. By contrast, the United Kingdom was a large exporter throughout the 1980s and the 1990s, but is now a net importer. Table A.1 in the Data Appendix provides more details on oil dependency across countries. Our strategy is to exploit this heterogeneity to more precisely estimate how economies respond to fluctuation in oil prices. Indeed, Figure 6 illustrates why a country’s oil dependence may go a long way in shaping the link between changes in oil prices and changes in consumption expenditures. First, we run country-specific OLS regressions of consumption (c ) on its lag, lagged real oil prices i,t (oil ), and world GDP (Y ): t−1 t c = ρ c +β (1−ρ )oil +η Y +u , (6) i,t i i,t−1 i i t−1 i t i,t where the country–specific coefficient β can be interpreted as the long–run elasticity of consumption to i changes in oil prices.7 If all countries were to respond similarly to changes in oil prices, one should not find a relationship between elasticities β ’s and countries’ oil dependence. Yet, as shown in Figure 6, the elasticity i of consumption to oil prices (β ’s in blue dots) is negatively correlated with oil dependency across countries. i Then, we propose a parsimonious approach to capture the heterogeneous responses of consumption to oil prices seen above. We estimate a panel regression in which the long-run elasticity of consumption to changes in oil prices, β˜ , depends on the country’s oil dependency, d , after a flexible transformation g(d ): i,t i,t i,t c = ρc +β˜ (1−ρ)oil +ηY +u , (7) i,t i,t−1 i,t t−1 t i,t β˜ = b +b g(d ;b ,b ). (8) i,t 0 1 i,t 2 3 Specifically, we assume that function g(d ) follows a strictly increasing Gompertz transformation: i,t g(d ;b ,b ) = exp{−exp[−b (d −b )]}, b > 0, b ∈ R. (9) i,t 2 3 2 i,t 3 2 3 By using this transformation, we can capture a wide range of relationships between an economy’s oil dependency, d , and its long-run elasticity of consumption to oil prices, β˜ : from a linear one to a nonlinear one i,t i,t in which lower oil dependency yields less than proportional increases in the elasticity to oil prices.8 Indeed, the red line of Figure 6 shows that as countries become less oil dependent, the elasticity of consumption to oil prices predicted by regression (7) increases by a magnitude less than proportional to what would be predicted by a simple linear relationship. 7We take logs of all variables in this regression and detrend them with a cubic trend, with the exact detrending matterring little for the results. Also, the actual regression has two lags on consumption. 8Weestimatethemodel(7)-(9)bynonlinearleastsquares. ConsistentwithFigure6,wealsoimposethatb >0andb <0. 0 1 7

3.2 Cross-Country Panel VAR Motivatedbythecross-countryheterogeneitydocumentedintheprevioussection,weestimateahybridpanel VAR that enriches the benchmark VAR of Section 2 by controlling for how the degree of oil dependency affects the macroeconomic response to an oil-price shock. The data used are the same as in Section 2: quarterly data for the 1975-2018 period, covering 55 countries, and totaling more than 6,700 observations. We specify the hybrid panel VAR as follows: y = γ y +γ Y +γ c +γ C +γ z i,t yy i,t−1 yY t−1 yc i,t−1 yC t−1 yz t−1 + (cid:0) γ +γ g(d ) (cid:1) oil +(δ +δ g(d ))noil +εy , (10) yo yd i,t−1 t−1 yo yd i,t−1 t−1 i,t c = α y +α Y +γ y +γ Y +γ c +γ C +γ z i,t cy i,t cY t cy i,t−1 cY t−1 cc i,t−1 cC t−1 cz t−1 +(γ +γ g(d ))oil +(δ +δ g(d ))noil +εc , (11) co cd i,t−1 t−1 co cd i,t−1 t−1 i,t z = α Y +α C +γ Y +γ C +γ z +γ oil +εz, (12) t zY t zC t zY t−1 zC t−1 zz t−1 zo t−1 t oil = α Y +α C +α z +γ Y +γ C +γ z +γ oil +εoil, (13) t oY t oC t oz t oY t−1 oC t−1 oz t−1 oo t−1 t noil ≡ max(0,oil −max(oil ,oil ,oil ,oil )), (14) t t t−1 t−2 t−3 t−4 where z is the Conference Board Composite Leading indicator (CLI) of business cycles conditions,9 oil is t t therealpriceofoil, c iscountry-specificconsumption, y iscountry-specificGDP,andC ≡ (cid:80)N ω ·c , i,t i,t t i=1 i,t i,t and Y ≡ (cid:80)N ω ·y are the corresponding world aggregates, with ω denoting country weights.10 All t i=1 i,t i,t i,t variables are measured in percent deviation from a quadratic trend. Note that this specification assumes that shocks to oil prices affect consumption, GDP, and commodity prices with a one period delay.11 The specification above departs from a standard panel VAR (Holtz-Eakin, Newey, and Rosen, 1988) in three ways. First, we allow countries’ responses of consumption and GDP to oil-price shocks to be a function of their time-varying oil dependence d . The specific functional form is the transformation i,t g(d ;b ,b ) estimated in the panel model (7)-(9). Second, we allow individual countries’ consumption and i,t 2 3 GDP to depend not only on their own lags, but also on lags of the world aggregates, thus modelling dynamic interdependencies across countries. The world aggregates are averages of the country-specific variables, as done for instance in the global VAR approach proposed by Pesaran (2006) and discussed in Canova and Ciccarelli (2013) and Chudik and Pesaran (2014). Third, we augment the panel VAR with the net real oil price increase variable (noil ), following Kilian and Vigfusson (2011) and similar to Hamilton (1996), thus t 9See Camacho and Perez-Quiros (2002) for CLI’s performance in anticipating business cycles conditions. The VAR results are similar if we use instaed the Conference Board Composite Coincident Indicator or global oil production. 10The estimated VAR has an additional lag for every lagged variable and intercepts. We omit these variables in equations (10)-(13) to save on notation. 11To see this, notice that only the lags of oil prices (oil ) and net oil price increases (noil ) are present in equations t−1 t−1 (10) and (11). However, differently from a standard VAR, the cross-equations restrictions embedded in the panel VAR (C ≡ t (cid:80)N ω ·c ,andY ≡ (cid:80)N ω ·y )produceerrorsthataregenerallycorrelatedacrossequations. SeeCanovaandCiccarelli i=1 i,t i,t t i=1 i,t i,t (2013) for discussion. 8

allowing for increases and decreases in oil prices to yield asymmetric effects on economic activity, following the lead of Davis and Haltiwanger (2001) and Hamilton (2003). Figure 7 plots the IRFs of consumption and GDP to oil-price shocks using the 2018 values of countries’ oil-dependencies. The choice of countries showcases the heterogeneous responses arising from these differing oil dependencies. When oil prices decline (Figure 7a), consumption and GDP of oil-importer countries, such as those in the euro area, increase but take time to fully materialize, reaching peaks after roughly 12 quarters. As in Section 2, consumption responds before GDP suggesting that wealth effects are larger than supply-side effects. Conversely, consumption and GDP of oil-exporter countries, such as Canada and Norway, quickly decrease, bottoming out 6 and 2 quarters after the shock, respectively. These results are consistent with a rapid tumble in investment, presumably in the energy sector, which then compounds with a protracted fall in consumption, possibly due to negative wealth effects. The United States, neither a large net oil-importer nor a large net oil-exporter in 2018, experiences more mixed effects: its GDP leans towards a contraction in the periods immediately after the shock, and only rises above trend many quarters later pulled by a modest rise in consumption. Figure 7 also shows that oil-price increases (Figure 7b) do more harm than the good afforded by oil price decreases (Figure 7a). In oil-importing economies, such as those in the euro area, the negative responses of GDP and consumption to oil price increases is larger in magnitude than the positive responses to oil price decreases, as in Hamilton (2003). In large oil-exporting economies, such as Norway, GDP and consumption experience short-lived expansions after a oil price increase, with these variables returning to their original levelafter6quarters. Thisresultcontrastswiththeprotracteddragtoeconomicactivityintheaftermathof oilpricedeclines. Onepossiblereasonforsuchasmallboosttooil-exportereconomiesfromoilpriceincreases is that these economies could suffer large drags from the downturn experienced by their oil-importer trading partners (which account for a much larger share of the world GDP). Table 1 summarizes these results, showing that the world GDP experiences only small boosts after oil price declines and much larger drags after oil price increases. InFigure8,weusetheestimatesofthecross-countrypanelVARtoquantifyhowthe“exogenous”portion of the 2014–2015 oil price slump affected consumption and GDP across countries. First and foremost, the VAR attributes a large chunk of the oil price decline throughout that period to surprise innovations to the oil price equation (εoil in equation (13)) in periods 2014Q3–2015Q1. Other shocks, as well as delayed t effects from other shocks taking place before 2014Q3, play a much more limited role in driving oil prices in 2014–2015. Accordingly, the response of oil prices—shown in the top left panel in terms of the 2014Q1 level— closely mirrors the actual path of oil prices. The remaining panels of Figure 8 show the response of consumption and GDP across countries. The top right panel shows the response for the world. Consumption rises gradually, peaking at about 0.8 percent above baseline in mid-2018. The response of GDP is initially slightly negative, likely reflecting some drag from declines in investment in the oil sector, but builds to also about 0.8 percent by the end of 2017. The middle panel compares the euro area and the United States. While the U.S. response mirrors that of the 9

world as a whole, the response of the euro area to the oil-price decline is larger than that of the United States, reflecting the eurozone’s larger oil dependence. In the bottom panel, GDP and consumption fall markedly in Canada and Norway, consistent with the fact that both economies are large oil exporters. InAppendixB,weconducttworobustnessexercises. First,wecomparetheoil-priceshocksεoil estimated t by our cross-country panel VAR (equations 10-14) with the oil-demand specific shocks identified by Kilian (2009). Figure B.1 shows that these shocks are similar, exhibiting a correlation of 0.49, which is higher than the one found by the analogous comparison done in Section 2.2. Second, we estimate our panel VAR starting in 1986, given a possible structural break in oil dynamics around this period (e.g., Baumeister and Peersman, 2013). The impulse responses for the restricted sample, shown in Figure B.2, are similar to those from our baseline model. 4 Oil Dependency Across U.S. States: State-Level Panel VAR In this section, we analyze how economic activity reacts to oil-price shocks across U.S. states by exploiting the cross-section and time series variation in states’ oil dependence. We show that consumption (proxied by car registrations) of states with higher oil-dependency increases more after exogenous oil price declines. We also show that changes in oil prices generate asymmetric effects: oil price decreases generate smaller boosts to aggregate U.S. consumption than the drag from oil price increases. 4.1 Oil Dependency Across the United States Just like countries in the world, U.S. states are very heterogeneous in their oil dependence.12 Figure 9 provides information about cross-sectional and time series variation in U.S. states’ oil dependency, using the measure defined in equation (5). Figure 9a shows that in 2013 states in the Northeast had an oil dependence of nearly 100 percent, while Texas was roughly oil independent. By contrast, states such as Alaska and North Dakota produced substantially more oil than consumed, with very large negative oil dependencies. Figure9bshowsthatU.S.statesalsoexhibitconsiderabletime-seriesvariationinoildependence, withNorth Dakota exporting increasing amounts of oil over time and Alaska exporting decreasing amounts. Table A.2 in the Appendix provides more details on oil dependency across U.S. states. In order to estimate the response of consumption to oil-price shocks at the state level, we use data on retail new car registrations as a proxy for consumption. Available at a quarterly frequency from 1989Q1 to 2018Q4, this dataset constructed by Polk’s National Vehicle Population Profile has important advantages over state-level data on personal consumption expenditures, as the latter are still experimental, are available only annually, and cover a shorter time frame. Appendix A.2 describes the state-level car registrations data and shows that retail registrations at the national level is highly correlated with NIPA expenditures on 12 In fact, using the cross-sectional standard deviation of oil dependency as a measure of heterogeneity, the U.S. states are actually more heterogeneous than the world’s countries in terms of oil dependency. In 2013, the cross-sectional standard deviation of oil dependency across U.S. states was 200. The corresponding number for the world’s countries was 130. 10

motor vehicles and parts. 4.2 State-Level Panel VAR We set up a panel VAR that mirrors the cross-country analysis discussed in the previous section. To do so, we use unemployment rates to measure overall economic activity at the state level. Thus, for each state i, let u denote the unemployment rate in quarter t and car the car registrations, with U ≡ (cid:80)N ω u i,t i,t t i=1 i,t i,t referring to the national unemployment, Car ≡ (cid:80)N ω car to the national car registrations, and ω t i=1 i,t i,t i,t to state-specific weights. The rest of the notation is the same as in Section 3: oil is the real price of oil, t noil is the real net oil price increase (equation (14)), and z is the Conference Board Composite Leading t t indicator (CLI). All variables are measured in percent deviation from a quadratic trend, with the exception of unemployment rates which are measured as percentage point deviation from a quadratic trend. With a specification analogous to the one of Section 3.2, the state-level panel VAR allows for heterogeneous regional responses to an oil shock as a function of oil dependence: u = γ u +γ U +γ car +γ Car +γ z i,t uu i,t−1 uU t−1 uc i,t−1 uC t−1 uz t−1 +(γ +γ g(d ))oil +(δ +δ g(d ))noil +εu , (15) uo ud i,t−1 t−1 uo ud i,t−1 t−1 i,t car = α u +α U +γ u +γ U +γ car +γ Car +γ z i,t cu i,t cU t cu i,t−1 cU t−1 cc i,t−1 cC t−1 cz t−1 +(γ +γ g(d ))oil +(δ +δ g(d ))noil +εc , (16) co cd i,t−1 t−1 co cd i,t−1 t−1 i,t z = α U +α Car +γ U +γ Car +γ z +γ oil +εz, (17) t zU t zC t zU t−1 zC t−1 zz t−1 zo t−1 t oil = α U +α Car +α z +γ U +γ Car +γ z +γ oil +εoil, (18) t oU t oC t oz t oU t−1 oC t−1 oz t−1 oo t−1 t where the functional form of g(d ) is described in equation (9) with coefficients estimated by a regression i,t analogous to (7).13 Figure 10a shows the predicted effect on car registrations of a 10 percent oil price decline triggered by an εoil shock. For the United States in aggregate, such shock is associated with a substantial increase in t car registrations, peaking at 1.3 percent a year after the shock, sustaining this increase for another year, and then experiencing fading effects. The large response of car registrations is consistent with the channels that emphasize the large oil price elasticity of demand for goods that are complementary with oil, and lines up well with the findings in the literature. For example, Edelstein and Kilian (2009) find that vehicle expenditures are highly sensitive to movements in the price of gasoline. Coglianese, Davis, Kilian, and Stock (2016) find an estimate of −0.37 for the price elasticity of gasoline demand, although their numbers do not appear directly comparable to ours, as they look at the cumulative effects of gasoline price changes. We then turn to the state-specific responses of car-registrations to oil price drops (Figure 10a). Oil- 13The estimated VAR has an additional lag for every lagged variable and intercepts. We omit these variables in equations (15)-(18) to save on notation. 11

importing states face effects similar to those seen for the United States in the aggregate, and thus are omitted from the figure. Oil-independent Texas, has a response similar to the U.S. aggregate, but with a magnitudeslightlysmaller. Bycontrast,majoroil-exportingstatessuchasNorthDakotasufferasubstantial decline in car registrations, with a trough of minus 1.3 percent in the quarter right after the shock. North Dakota’s car registrations soon rise but only to a level not statistically different from zero. This rise may reflect the fact that, because cars run on gas, lower oil prices make car-buying more attractive, even in oil-intensive states. Moreover, the subsequent return of car registrations to North Dakota’s original level may also reflect the increase in overall U.S. aggregate demand, with the close trade and financial linkages within the U.S. states dampening the effects of heterogeneity in oil production. Figure 10b then shows that the benefits from oil-price decreases are smaller than the losses from oilprice increases also across U.S. states. We see that shocks pushing up oil prices are followed by a national decreaseincarregistrations. Thisdecreaseislargerthantheincreaseinregistrationsfollowingacomparable oilpricedrop(Figure10a). TheresultsforTexasarequalitativelysimilartothoseforthenationalaggregate, although with smaller magnitudes. In the case of oil-rich North Dakota, car registrations also fall after an oil-price increase, possibly due to the U.S. aggregate demand effect. The Appendix with Supplementary Results (Figure B.4) shows that results for unemployment rates are similar to those for car registrations, with the main difference being that unemployment rates react a bit more slowly to oil-price shocks than car registrations. All told, the state-level evidence confirms the findings of the country-level VAR, highlighting once again howvaryingoildependencymayimplydifferencesinconsumptionresponsesacrossregions, whereonewould think that common monetary and fiscal policy implies somewhat similar regional responses. However, as is well known in the literature on fiscal and monetary unions, a common policy response may either amplify or dampentheheterogeneityofregionalresponses, especiallywhenshockspresentauthoritieswithstabilization trade-offs. Consider for instance an oil shock that lowers inflation and activity in North Dakota, but lowers inflation and boost activity in Connecticut. If the policy goal is to stabilize activity, such shock calls for expansionarypoliciesinNorthDakota, andforcontractionarypoliciesinConnecticut. Heterogeneouspolicy responseswouldthenimplyasimilarresponseofactivityinNorthDakotaandConnecticutthantheresponse implied by a common policy response. 5 Literature Review Thereisavastliteraturethatanalyzeshowfluctuationsinoilpricesaffectmacroeconomicoutcomes. Inparticular, our paper belongs to a growing set of studies that investigate the responses to oil-price shocks across different economies. These studies have explored variation in geographical location, institutional development, and net-exporting position of energy across different countries and states of the Unites States (e.g., Aastveit,Bjrnland,andThorsrud,2015,Guerrero-Escobar,Hernandez-delValle,andHernandez-Vega,2018, Peersman and Van Robays, 2012, and Bjørnland and Zhulanova, 2018). In addition, Baumeister, Peersman, 12

Van Robays, et al., 2010 (henceforth abbreviated as BPVR) is particularly related to our work because of its cross-country perspective. BPVR estimate country-specific VARs for eight different countries and document that net-energy importing economies suffer falls in economic activity following an oil-supply shock, with these effects being not significant or positive for net-energy exporters. BPVR also show qualitative evidence that improvements in the net-energy position of countries is associated with more muted responses to oil-supply shocks. Our paper makes four contributions. First, we explore a much larger set of economies than previous papers by using two panel datasets: one with 55 countries and another with all U.S. states. Second, we exploit the variation in oil dependency both across countries and over time—beyond the status of oilimporter/exporter—thus better quantifying how changes in oil-dependency impact an economy’s response to oil-price shocks. Third, we measure the effect of oil shocks on aggregate consumption, a variable often overlooked in the literature. We show that oil price declines generate large and positive effects on consumption and economic activity (GDP and unemployment rate) in oil-importing economies, while depressing consumption and activity in oil-exporting economies. Fourth, we allow asymmetric responses to oil-price shocks, with macroeconomic effects from oil price increases being potentially different from oil-price decreases. We then show that oil prices increases do more harm than the good afforded by oil-price decreases across both the world and the United States. 6 Conclusions In this paper, we provide empirical evidence that lower oil prices boost consumption in oil importing countries, and depress it in oil exporting ones. This effect varies in proportion to the degree of oil dependency on an economy. For instance, a 10 percent decline in oil prices boosts euro area consumption and GDP by about 0.2 percent, while leading to declines of similar magnitudes in consumption and GDP in oil-exporting country such as Canada. In between, there are countries like the United States where the effects are more mixed. While in the short run U.S. GDP might temporarily decline, presumably reflecting weaker investment in the oil industry, a gradual rise in U.S. consumption pushes its GDP towards an modest expansion. For the world aggregate, oil price declines boost economic activity, although by smaller magnitudes than similar increases in oil prices. We complement our cross-country results by showing analogous findings for the U.S. states and aggregate. Our results have important implications for the design of monetary policy in the presence of oil price fluctuations. A large and growing literature documents how optimal monetary policy may entail different responses to oil price increases depending on the source of the shock, say an increase in foreign demand or a glut in foreign supply. For instance, Bodenstein, Guerrieri, and Kilian (2012) develop and estimate a DSGE model with oil in which multiple shocks drive oil price fluctuations. In their model, the peak response of output and consumption to oil shocks happens in the same period of the shock. They find that no two shocksinducethesamepolicyresponse,evencontrollingforthesamepathofoilpricesimpliedbytheshocks. 13

They further show that, in the wake of fluctuations in oil prices, a monetary policy rule that places a higher weight on stabilizing wage inflation fosters the stabilization of core inflation. Our evidence reinforces their argument, by showing that the same shock may imply different macroeconomic responses depending on the oil dependency of a particular country. It also illustrates an important challenge for extant models of the oil market and the macroeconomy, which may want to incorporate features such as habits and adjustment costs in order to capture the delays in which oil shocks affect economic activity. Our evidence shows that changes in oil dependence may imply non-trivial effects of changes in oil prices, that the effects of oil shocks on consumption may take time to materialize, and that the response of GDP to an oil shock is not equal to the mere contribution of consumption, presumably because investment in the energy sector may be sensitive to changes in oil prices too. We view such evidence, gathered from a large set of countries over a long time period, as providing a useful empirical benchmark to gauge the plausibility of the policy recommendations of optimizing monetary models of the business cycle that incorporate an oil sector. 14

References Aastveit, K. A., H. C. Bjrnland, and L. A. Thorsrud (2015): “What Drives Oil Prices? Emerging Versus Developed Economies,” Journal of Applied Econometrics, 30(7), 1013–1028. 2, 12 Arezki, R., and O. Blanchard (2014): “Seven questions about the recent oil price slump,” IMF Blog. 2 Baumeister, C., and G. Peersman (2013): “The Role Of Time-Varying Price Elasticities In Accounting For Volatility Changes In The Crude Oil Market,” Journal of Applied Econometrics, 28(7), 1087–1109. 10 Baumeister, C., G. Peersman, I. Van Robays, et al. (2010): “The economic consequences of oil shocks: differences across countries and time,” Inflation in an era of relative price shocks, Reserve Bank of Australia, pp. 91–128. 12 Bernanke, B., J. Boivin, and P. Eliasz (2005): “Factor augmented vector autoregressions (FVARs) and the analysis of monetary policy,” Quarterly Journal of Economics, 120(1), 387–422. 4 Bernanke, B. S. (2016): “The relationship between stocks and oil prices,” Brookings Institution. 2 Bjørnland, H. C., and J. Zhulanova (2018): “The Shale Oil Boom and the US Economy: Spillovers and Time-Varying Effects,” BI Norwegian Business School Working Paper. 12 Bodenstein, M., C. J. Erceg, and L. Guerrieri(2011): “Oilshocksandexternaladjustment,” Journal of International Economics, 83(2), 168–184. 3 Bodenstein, M., L. Guerrieri, and L. Kilian (2012): “Monetary Policy Responses to Oil Price Fluctuations,” IMF Economic Review, 60(4), 470–504. 13 Caldara, D., M. Cavallo, and M. Iacoviello (2019): “Oil price elasticities and oil price fluctuations,” Journal of Monetary Economics, 103, 1 – 20. 5 Camacho, M., and G. Perez-Quiros (2002): “This is what the leading indicators lead,” Journal of Applied Econometrics, 17(1), 61–80. 8 Canova, F., and M. Ciccarelli (2013): Panel Vector Autoregressive Models: A Surveychap. 12, pp. 205–246. 8 CBS/AP(29Feb.2016): “BuffettSaysU.S.EconomyWeakerthanHeExpectedbutGrowing,”CBS Money Watch, Web. 2 Chudik, A., and M. H. Pesaran(2014): “TheoryandpracticeofGVARmodelling,”Journal of Economic Surveys. 8 Coglianese, J., L. W. Davis, L. Kilian, and J. H. Stock (2016): “Anticipation, tax avoidance, and the price elasticity of gasoline demand,” Journal of Applied Econometrics. 11 15

Davis, S. J., and J. Haltiwanger (2001): “Sectoral job creation and destruction responses to oil price changes,” Journal of monetary economics, 48(3), 465–512. 9 Edelstein, P., and L. Kilian (2009): “How sensitive are consumer expenditures to retail energy prices?,” Journal of Monetary Economics, 56(6), 766–779. 11 Guerrero-Escobar, S., G. Hernandez-del Valle, and M. Hernandez-Vega (2018): “Do heterogeneous countries respond differently to oil price shocks?,” Journal of Commodity Markets. 12 Hamilton, J. D. (1996): “This is what happened to the oil price-macroeconomy relationship,” Journal of Monetary Economics, 38(2), 215–220. 8 (2003): “What is an oil shock?,” Journal of econometrics, 113(2), 363–398. 9 (2008): “Oil and the Macroeconomy,” The New Palgrave Dictionary of Economics. 3 Holtz-Eakin, D., W. Newey, and H. S. Rosen (1988): “Estimating vector autoregressions with panel data,” Econometrica, pp. 1371–1395. 8 Jime´nez-Rodr´ıguez, R., and M. Sa´nchez (2005): “Oil price shocks and real GDP growth: empirical evidence for some OECD countries,” Applied economics, 37(2), 201–228. 2 Kilian, L. (2009): “Not All Oil Price Shocks Are Alike: Disentangling Demand and Supply Shocks in the Crude Oil Market,” American Economic Review, 99(3), 1053–69. 5, 6, 10, 20, B.1 Kilian, L., A. Rebucci, and N. Spatafora (2009): “Oil shocks and external balances,” Journal of International Economics, 77(2), 181–194. 2 Kilian, L., and R. J. Vigfusson (2011): “Are the responses of the U.S. economy asymmetric in energy price increases and decreases?,” Quantitative Economics, 2(3), 419–453. 8 Peersman, G., and I. Van Robays (2012): “Cross-country differences in the effects of oil shocks,” Energy Economics, 34(5), 1532–1547. 12 Pesaran, M. H. (2006): “Estimation and inference in large heterogeneous panels with a multifactor error structure,” Econometrica, 74(4), 967–1012. 8 Stock, J. H., and M. W. Watson (2005): “Implications of Dynamic Factor Models for VAR Analysis,” NBER Working Papers 11467, National Bureau of Economic Research, Inc. 4 Stock, J. H., and M. W. Watson (2016): “Dynamic factor models, factor-augmented vector autoregressions, and structural vector autoregressions in macroeconomics,” in Handbook of macroeconomics, vol. 2, pp. 415–525. Elsevier. 5 16

Table 1: Responses of Consumption and GDP in Cross-Country Panel VAR 10% Oil Price Decline Euro Area U.S. Canada Norway World Consumption 0.16 0.06 -0.13 -0.40 0.07 GDP 0.10 0.06 -0.04 -0.17 0.06 10% Oil Price Increase Euro Area U.S. Canada Norway World Consumption -0.23 -0.16 -0.02 0.16 -0.17 GDP -0.16 -0.14 -0.09 -0.03 -0.14 Note: TheconsumptionandGDPresponsesaremeasuredastheaverageoftheirimpulseresponsefunctions (IRFs) during the first 3 years (from quarter 1 through quarter 12) after a shock moving oil prices by 10 percent (in quarter 0). These IRFs are calculated using the cross-country panel VAR from Section 3.2. 17

Figure 1: Oil Prices and Consumption Differential Correlation = -0.32*** Consumption Differential Real Oil Price egnahc tnecrep retrauq-4 fo ecnereffid tpP 01 5 0 5- 051 001 05 0 05egnahc tnecrep retrauq-4 1975q1 1982q1 1989q1 1996q1 2003q1 2010q1 2018q4 Note: Consumption differential is the aggregate consumption growth of oil importers minus the aggregate consumption growth of oil exporters (red line with y-axis on the left). Real oil prices (green line with yaxis on the right) are measured in 2009 dollars. While consumption differential is measured in percentage point difference in 4-quarter percent changes, real oil prices are measured in 4-quarter percent change. The correlation between consumption differential and real oil prices is statistically significant at the 99 percent confidence level (***). Exporters and importers are aggregated using weights based on each country’s GDP in constant US dollars. The list of countries and their classification of importer/exporter is in Table A.1. 18

Figure 2: Impulse Response Functions to Oil-Price Shock in Aggregate VAR Oil Price Cons. Importers Cons. Exporters 5 0.3 0.3 0.2 0.2 0 0.1 0.1 tn tn tn e e e c c 0 c 0 r r r e e e P -5 P P -0.1 -0.1 -0.2 -0.2 -10 -0.3 -0.3 0 10 20 0 10 20 0 10 20 Oil Production GDP Importers GDP Exporters 0.6 0.3 0.3 0.2 0.2 0.4 0.1 0.1 tn tn tn e e e c 0.2 c 0 c 0 r r r e e e P P P -0.1 -0.1 0 -0.2 -0.2 -0.2 -0.3 -0.3 0 10 20 0 10 20 0 10 20 Quarters Quarters Quarters Note: The figure shows the impulse response functions (IRFs) to oil-price shocks in the aggregate VAR (Section 2.1). Solid lines represent median IRFs. Shaded areas represent 16th and 84th percentiles constructed with 5,000 bootstrap replications. Variables are expressed as percent deviation from quadratic trend. 19

Figure 3: Comparison of Oil Shocks Identified under Different Strategies 5 Oil-Price Shocks (Aggregate VAR) 4 KOD Shocks Correlation=0.41*** 3 s 2 n o ita iv 1 e D d r 0 a d n a tS -1 -2 -3 -4 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 Note: “Oil-Price Shocks (Aggregate VAR)” represent the oil shocks identified by the aggregate VAR from Section 2.1 of this paper. “KOD shocks” represent the “oil-specific demand” shocks identified by Kilian (2009) averaged to the quarterly frequency. The correlation between the two series of shocks is statistically significant at the 99 percent confidence level (***). 20

Figure 4: Impulse Response Functions to Oil Shock under Alternative Oil-Supply Elasticity Oil Price Cons. Importers Cons. Exporters 5 0.25 0.2 0 0 tn tn tn e e e c -5 c c r e r e0.1 r e P P P -0.25 -10 Alternative Elasticity Benchmark -15 0 -0.5 0 10 20 0 10 20 0 10 20 Oil Production GDP Importers GDP Exporters 0.2 2 0.2 1.5 0 tn tn tn e e0.1 e c 1 c c r r r e e e P P P -0.2 0.5 0 0 -0.4 0 10 20 0 10 20 0 10 20 Quarters Quarters Quarters Note: “Benchmark” impulse response functions (IRFs) are calculated using the aggregate VAR of Section 2.1andareshowningreen. “AlternativeElasticity”IRFs,showninblue,arecalculatedundertheassumption of a world oil demand elasticity of -0.21 and world supply elasticity of 0.1, as explained in Section 2.2. Solid lines represent median IRFs. Shaded areas represent 16th and 84th percentiles constructed using 5,000 bootstrap replications. Variables are expressed as percent deviation from quadratic trend. 21

Figure 5: Oil Dependency Across Countries and Over Time (a) Oil Dependency (in Percent) Across Countries in 2013 Norway Russia Venezuela Colombia Saudi Arabia Ecuador Iran Canada Mexico Denmark Brazil United Kingdom United States Indonesia Australia India Italy Germany Netherlands Korea France Japan Spain Switzerland Turkey -800 -600 -400 -200 0 100 mean of oildep1_ma_pct (b) Oil Dependency (in Percent) Across a Select Group of Countries Over Time Brazil United States Canada United Kingdom Mexico )tnecreP ni( ycnednepeD liO 001 05 0 05- 001- 1975q1 1982q1 1989q1 1996q1 2003q1 2010q1 2018q4 Note:Thetoppanelshowsoildependencyin2013forasubsetofthelargestcountriesinthesample(largest exporters and countries with share of world GDP in 2013 larger than 1 percent), where oil dependency is measured by oil imports as a percentage of oil consumption (equation 5). The bottom panel plots oil dependency over time for a select group of economies. 22

Figure 6: Oil Dependency and Long-Run Elasticity of Consumption to Oil Prices 0.12 South Africa s e c ir Venezuela P liO 0.06 Russia o t n Norway Ecuador Pe r B u razil o itp Malaysia m Argentina Colombia u s Saudi Arabia Iran n 0 Canada France o C Spain Mexico fo Italy y tic its a Germany lE Indonesia n -0.06 United Kingdom u R - New Zealand g n o L Korea Country-Specific Regressions -0.12 Panel Regression United States -800 -700 -600 -500 -400 -300 -200 -100 0 100 200 300 Oil Dependency (in Percent) Note: On the horizontal axis, oil dependency is calculated as the sample average, where oil dependency is measured by oil imports as a percentage of oil consumption (equation 5). The vertical axis features long-run elasticities of consumption to changes in oil prices. These elasticities are estimated either by country-bycountry regressions (6) with fitted β ’s represented by blue dots, or by panel regression (7) with fitted values i represented by the red line. 23

Figure 7: Impulse Response Functions to Oil-Price Shocks in Cross-Country Panel VAR (a) Oil Price Decrease (b) Oil Price Increase Oil Prices Oil Prices 5 10 0 tn e tn e 5 c re -5 c re P P 0 -10 -5 0 5 10 15 20 0 5 10 15 20 Euro Area Euro Area 0.4 0.4 0.2 0.2 tn tn e e c re 0 c re 0 P P -0.2 -0.2 -0.4 -0.4 0 5 10 15 20 0 5 10 15 20 United States United States 0.4 0.4 0.2 0.2 tn tn e e c re 0 c re 0 P P -0.2 -0.2 -0.4 -0.4 0 5 10 15 20 0 5 10 15 20 Canada Canada 0.4 0.4 0.2 0.2 tn tn e e c re 0 c re 0 P P -0.2 -0.2 -0.4 -0.4 0 5 10 15 20 0 5 10 15 20 Norway Norway 0.75 0.75 0.5 0.5 tn 0.25 tn 0.25 e e c re 0 c re 0 P-0.25 P-0.25 -0.5 -0.5 -0.75 -0.75 0 5 10 15 20 0 5 10 15 20 Quarters Quarters Consumption GDP Consumption GDP Note: The figure shows the impulse response functions (IRFs) to oil-price shocks in the cross-country panel VAR (Section 3.2). Solid lines represent median IRFs. Shaded areas represent 16th and 84th percentiles constructed with 5,000 bootstrap replications. Variables are expressed as percent deviation from quadratic trend. 24

Figure 8: Contribution of the 2014-2015 Oil-Price Slump to GDP and Consumption Oil Prices World 100 le 0.8 v e 0 L 0 1 90 1 0.6 = Q 1 4 Q 4 80 1 0 0.4 1 2 0 m 2 0.2 :x e d 70 o r f tn 0 n e I 60 c r e P -0.2 2014 2015 2016 2017 2018 2019 2020 2014 2015 2016 2017 2018 2019 2020 Euro Area United States 0.8 le le v v e e L L 0.6 1 1 Q 1 Q 4 4 0.4 1 1 0 0 2 2 m m 0.2 o 0.5 o r r f tn f tn 0 e e c c r r e e -0.2 P 0 P 2014 2015 2016 2017 2018 2019 2020 2014 2015 2016 2017 2018 2019 2020 Canada Norway 0.2 0 le le v v e L 0 e L -0.5 1 1 Q Q 4 -0.2 4 -1 1 1 0 0 2 -0.4 2 -1.5 m m o o r f tn -0.6 r f tn -2 e e c -0.8 c -2.5 r r e e P P -1 2014 2015 2016 2017 2018 2019 2020 2014 2015 2016 2017 2018 2019 2020 Consumption GDP Note: The figure show the historical contribution of oil-price shocks to consumption and GDP of the world aggregate and selected countries. Oil-price shocks are those estimated for the periods 2014Q3 through 2015Q1 using the cross-country panel VAR (Section 3.2). Variables are expressed as percent deviation from the 2014Q1 level, with the exception of oil prices which is measured as an index with level 100 at 2014Q1. 25

Figure 9: Oil Dependency Across U.S. States and Over Time (a) Oil Dependency (in Percent) Across U.S. States in 2013 North Dakota = -1100 Alaska New Mexico Wyoming Oklahoma Montana Colorado Texas Kansas Utah California Louisiana Alabama Arkansas Michigan Illinois Ohio Pennsylvania Kentucky Indiana Florida Tennessee Missouri New York Arizona Virginia Connecticut Georgia Maryland Massachusetts Minnesota New Jersey North Carolina Oregon South Carolina Washington Wisconsin -800 -600 -400 -200 0 100 (b) Oil Dependency (in Percent) Across a Select Group of U.S. States Over Time California Texas Wyoming Alaska North Dakota Alaska oil dep.< -1500 before 1998 North Dakota oil dep.< -1500 after 2015 )tnecreP ni( ycnednepeD liO 0010 005- 0001- 0051- 1990q1 2000q1 2010q1 2020q1 Note:Thetoppanelshowsoildependencyin2013forasubsetofthelargeststatesintheU.S.(oilproducers and states with share of car purchases in 2013 larger than 1 percent), where oil dependency is measured by oil imports as a percentage of oil consumption (equation 5). The bottom panel plots oil dependency over time for a select group of U.S. states. 26

Figure 10: Impulse Response Functions to Oil-Price Shocks in State-Level Panel VAR (a) Oil Price Decrease (b) Oil Price Increase OOili lP Prirciceess OOili lP Prirciceess 1010 1010 55 55 tn tn tn tn ee ee cc 00 cc 00 r e r e r e r e PP PP -5-5 -5-5 -1-010 -1-010 00 55 1010 1515 2020 00 55 1010 1515 2020 CCaar rR Reeggisistrtaratitoionn: :U U.S.S. .A Aggggrereggaatete CCaar rR Reeggisistrtaratitoionn: :U U.S.S. .A Aggggrereggaatete 33 33 1.15.5 1.15.5 tn tn tn tn ee ee cc 00 cc 00 r e r e r e r e PP PP -1-.15.5 -1-.15.5 -3-3 -3-3 00 55 1010 1515 2020 00 55 1010 1515 2020 CCaar rR Reeggisistrtaratitoionn: :T Teexxaass CCaar rR Reeggisistrtaratitoionn: :T Teexxaass 33 33 1.15.5 1.15.5 tn tn tn tn ee ee cc 00 cc 00 r e r e r e r e PP PP -1-.15.5 -1-.15.5 -3-3 -3-3 00 55 1010 1515 2020 00 55 1010 1515 2020 CCaar rR Reeggisistrtaratitoionn: :N Noortrhth D Daakkootata CCaar rR Reeggisistrtaratitoionn: :N Noortrhth D Daakkootata 33 33 1.15.5 1.15.5 tn tn tn tn ee ee cc 00 cc 00 r e r e r e r e PP PP -1-.15.5 -1-.15.5 -3-3 -3-3 00 55 1010 1515 2020 00 55 1010 1515 2020 QQuuaartretersrs QQuuaartretersrs Note: The figure shows the impulse response functions (IRFs) to oil-price shocks in the U.S. state-level panel VAR (Section 4). Solid lines represent median IRFs. Shaded areas represent 16th and 84th percentiles constructed using 5,000 bootstrap replications. Variables are expressed as percent deviation from quadratic trend. 27

Appendix A The Data Appendix A.1 Country-Level Data The sample includes 55 countries for which it was possible to assemble sufficiently long and reliable data on quarterly real consumption and quarterly real GDP from the country’s official statistics. Of the richest countries in the world (measured by total GDP in US dollars), we do not include China because of the lack of consumption data at a quarterly frequency. By the same reason, we exclude some of the largest oil producers, such as United Arab Emirates, Iraq, Kuwait, Nigeria, Qatar, Angola, Algeria, and Kazakhstan. In the aggregate VAR analysis, we construct GDP and consumption using data for real consumption and real GDP using weights based on each country’s GDP in constant U.S. dollars from the World Bank World Development Indicators (if the data are missing, the weight is set at zero, and the weights of other countries are changed accordingly). We list in A.1 all countries in our sample, their oil dependency, their sample weights for the year 2013, and their data coverage. Oil dependency is constructed using annual data on oil production and consumption from the BP Statistical Review of World Energy (with data dating back to 1965). An advantage of these data is that, barring small changes in year-to-year inventories, they satisfy the identity that oil production must be equal oil consumption over very long horizons. For countries that are small oil producers, the BP Statistical Review does not report data (shown as missing in Table A.1). In these cases, we search for oil production and consumption in the CIA World Factbook and the International Energy Agency (IEA). If we find these oil statistics, we set oil dependency to a constant implied by the data for all years. If we are not able to find the data in either the BP Statistical review, the CIA World Factbook or the IEA, we set oil dependency to 1 in all years. In all these cases of small oil producers, the implied oil dependency is above 0.65. Appendix A.2 State-Level Data The key variable included in our state-level panel vector autoregression is “new car registrations, retail” broken down by state of registration. This measure is constructed using data from Polk’s National Vehicle Population Profile (NVPP), which is a census of all currently registered passenger cars and light-duty trucks in the United States. Polk’s New Registration Data provides detailed indicators for new vehicle registrations(includingwherethevehiclewasregistered). The“Retail”filerepresentsnewvehiclespurchased by individuals, thus excluding registrations of rental, fleet, government and other commercial use vehicles. From 1989 to 2002, Polk’s data is computed by marketing regions, with information also available for states. We then splice this dataset with the one for 2002 onwards, which is computed directly at the state level. As shown by Figure B.3, retail registrations at the national level are very highly correlated with durable goods, namely Motor vehicles and parts from the National Income and Product accounts. A.1

Table A.1: Cross-Country Oil Dependency in 2013 and Data Availability Country GDP Share Oil Oil Oil Exporter Start End Dependency Consumption Production Date Date UnitedStates 27.03 47 18961 10071 0 1975q1 2018q4 Japan 10.05 97 4516 0 1975q1 2018q4 Germany 6.10 93 2408 0 1975q1 2018q4 France 4.65 97 1664 0 1975q1 2018q4 UnitedKingdom 4.40 43 1518 864 0 1975q1 2018q4 Brazil 4.11 32 3124 2110 0 1990q1 2018q4 Italy 3.48 91 1260 114 0 1975q1 2018q4 India 3.37 76 3727 906 0 1996q2 2018q4 Canada 2.96 -68 2383 4000 1 1975q1 2018q4 Russia 2.89 -245 3135 10809 1 1995q1 2018q4 Spain 2.31 97 1195 0 1975q1 2018q4 Australia 2.13 61 1034 407 0 1975q1 2018q4 Korea 2.04 96 2455 0 1975q1 2018q4 Mexico 1.96 -41 2034 2875 1 1980q1 2018q4 Turkey 1.66 100 757 0 1987q1 2018q4 Indonesia 1.53 47 1663 882 0 1983q1 2018q4 Netherlands 1.45 94 898 0 1975q1 2018q4 SaudiArabia 1.07 -230 3451 11393 1 2005q1 2018q4 Switzerland 1.04 100 249 0 1975q1 2018q4 Poland 0.88 93 538 0 1995q1 2018q4 Sweden 0.86 96 309 0 1975q1 2018q4 Belgium 0.84 98 636 0 1975q1 2018q4 Taiwan 0.83 100 1010 0 1975q1 2018q4 Iran 0.79 -80 2011 3617 1 2004q2 2018q2 Argentina 0.78 5 683 647 0 1993q1 2018q4 Norway 0.77 -656 243 1838 1 1975q1 2018q4 Venezuela 0.75 -243 782 2680 1 1997q1 2015q4 Austria 0.69 87 264 0 1975q1 2018q4 SouthAfrica 0.69 73 572 0 1975q1 2018q4 Thailand 0.65 65 1299 452 0 1993q1 2018q4 Colombia 0.57 -237 298 1004 1 2000q1 2018q4 Denmark 0.56 -13 158 178 1 1975q1 2018q4 Malaysia 0.51 22 803 626 0 1991q1 2018q4 Singapore 0.47 98 1225 0 1975q1 2018q4 Israel 0.44 98 247 0 1995q1 2018q4 Chile 0.43 95 362 0 1996q1 2018q4 HongKong 0.43 100 311 0 1990q1 2018q4 Finland 0.42 95 191 0 1975q1 2018q4 Greece 0.42 98 295 0 1975q1 2018q4 Philippines 0.40 94 326 0 1981q1 2018q4 Ireland 0.40 99 138 0 1975q1 2018q4 Portugal 0.38 97 239 0 1975q1 2018q4 CzechRepublic 0.36 94 184 0 1996q1 2018q4 Peru 0.30 26 227 167 0 1980q1 2018q4 NewZealand 0.27 66 151 0 1975q1 2018q4 Hungary 0.23 82 129 0 1995q1 2018q4 Slovakia 0.16 89 75 0 1995q1 2018q4 Ecuador 0.14 -113 247 527 1 1990q1 2018q4 Luxembourg 0.10 100 57 0 1975q1 2018q4 Slovenia 0.08 99 50 0 1995q1 2018q4 Latvia 0.05 97 33 0 1995q1 2018q4 Estonia 0.04 65 31 0 1995q1 2018q4 ElSalvador 0.03 100 25 0 1990q1 2018q4 Botswana 0.03 100 17 0 1994q1 2018q4 Iceland 0.03 100 15 0 1997q1 2018q4 Note: Oil dependency is calculated as oil consumed minus oil produced as a percentage of oil consumed (source: BP Statistical Review of World Energy downloaded in June 2019). Oil production is production of crude oil, tight oil, oil sands, and NGLs (thousand barrels daily). Oil consumption is inland demand plus international aviation and marine bunkers and refinery fuel and loss (thousand barrels daily). In the aggregate VAR, the “exporters” are countries with negative oil dependency in 2013. A.2

Table A.2: United States State-Level Data in 2013 State Car Share Oil Oil Oil Production Dependency Consumption California 11.45 48 1713 891 Texas 8.55 -16 3564 4153 Florida 7.14 99 824 10 NewYork 5.77 100 658 2 Pennsylvania 4.45 96 618 24 Illinois 4.26 93 627 43 Michigan 4.21 93 449 32 Ohio 4.17 94 589 36 NewJersey 3.55 100 516 0 Georgia 2.94 100 471 0 NorthCarolina 2.86 100 437 0 Virginia 2.58 100 425 0 Massachusetts 2.25 100 298 0 Maryland 2.17 100 253 0 Arizona 1.88 100 267 0 Indiana 1.86 97 398 11 Missouri 1.86 100 325 1 Tennessee 1.84 100 352 1 Wisconsin 1.81 100 279 0 Washington 1.71 100 369 0 Louisiana 1.62 71 1102 322 Colorado 1.48 -21 246 297 Alabama 1.48 83 268 47 Minnesota 1.47 100 318 0 SouthCarolina 1.36 100 262 0 Connecticut 1.27 100 167 0 Oklahoma 1.14 -93 267 516 Kentucky 1.10 97 303 10 Oregon 1.09 100 172 0 Arkansas 0.94 83 172 30 Iowa 0.92 100 241 0 Kansas 0.79 -16 180 210 Mississippi 0.79 50 219 109 Nevada 0.77 99 119 1 WestVirginia 0.65 69 106 32 Utah 0.63 -6 148 157 NewHampshire 0.60 100 80 0 NewMexico 0.60 -275 123 460 Nebraska 0.55 90 123 13 Maine 0.44 100 99 0 Hawaii 0.40 100 116 0 Delaware 0.36 100 51 0 Idaho 0.36 100 84 0 RhodeIsland 0.33 100 44 0 Vermont 0.28 100 44 0 Montana 0.28 -47 89 131 SouthDakota 0.24 86 61 8 NorthDakota 0.22 -1119 115 1398 Alaska 0.20 -591 122 841 Wyoming 0.19 -246 82 284 DistrictOfColumbia 0.14 100 10 0 Note: Oil dependency is calculated as oil consumed minus oil produced as a percentage of oil consumed (source: EIA: State Profiles and Energy Estimates). Oil production is production of crude oil by state (sumcrudestate). Production of crude oil by state is scaled up by the ratio of total U.S. oil production from the BP Statistical Review (which includes tight oil, oil sands, and other NGLs and including offshore production) to sumcrudestate (which is about 1.5), so as to facilitate comparison with oil dependency measures by country. Oil consumption includes all petroleum products consumed, and lines up with oil consumption from BP Statistical Review. Production and consumption are in thousand of barrels daily. A.3

Appendix B Supplementary Results Figure B.1: Comparison of Oil Shocks Identified under Different Strategies 5 Oil-Price Shocks (Cross-Country Panel VAR) 4 KOD Shocks Correlation=0.49*** 3 s 2 n o ita iv 1 e D d r 0 a d n a tS -1 -2 -3 -4 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 Note: “Oil-Price Shocks (Cross-Country Panel VAR)” represent the oil shocks identified by the panel VAR fromSection3.2ofthispaper. “KODshocks”representthe“oil-specificdemand”shocksidentifiedbyKilian (2009) averaged to the quarterly frequency. The correlation between the two series of shocks is statistically significant at the 99 percent confidence level (***). B.1

Figure B.2: Impulse Response Functions to Oil-Price Shocks in Cross-Country Panel VAR, Data Post-1986 Oil Prices World 0.3 0.2 0 0.1 tn tn e e 0 c c r -5 r e e P P -0.1 -0.2 -10 -0.3 0 5 10 15 20 0 5 10 15 20 USA Euro Area 0.3 0.3 0.2 0.2 0.1 0.1 tn tn e 0 e 0 c c r r e e P -0.1 P -0.1 -0.2 -0.2 -0.3 -0.3 0 5 10 15 20 0 5 10 15 20 Canada Norway 0.3 0.4 0.2 0.2 0.1 tn tn e 0 e 0 c c r r e e P -0.1 P -0.2 -0.2 -0.4 -0.3 -0.6 0 5 10 15 20 0 5 10 15 20 Quarters Consumption GDP Quarters Note: The figure shows the impulse response functions (IRFs) to oil-price shocks in the cross-country panel VAR(Section3.2)usingdatafrom1986onwards. SolidlinesrepresentmedianIRFs. Shadedareasrepresent 16th and 84th percentiles constructed with 5,000 bootstrap replications. Variables are expressed as percent deviation from quadratic trend. B.2

Figure B.3: New Cars Registrations and Motor Vehicles Expenditures from NIPA Correlation = 0.85 raey-no-raey ,egnahc % 02 0 02- 04- 1990q1 2000q1 2010q1 2020q1 New Car Registrations PCE: Motor Vehicles and Parts Note: This figures compares NIPA expenditures on motor vehicle (in real terms) with the aggregate data on new cars’ registrations from Polks National Vehicle Population Profile. B.3

Figure B.4: Impulse Response Functions to Oil-Price Shocks in State-Level Panel VAR (a) Oil Price Decrease (b) Oil Price Increase OOili lP Prriciceess OOili lP Prriciceess 1100 1100 55 55 tntn tntn ee ee cc 00 cc 00 r er e r er e PP PP -5-5 -5-5 -1-100 -1-100 00 55 1100 1155 2200 00 55 1100 1155 2200 UUnneemmpploloyymmeenntt R Raattee:: U U.S.S. .A Aggggrreeggaattee UUnneemmpploloyymmeenntt R Raattee:: U U.S.S. .A Aggggrreeggaattee 00.3.3 00.3.3 ss ss tn io tn io 00.1.155 tn io tn io 00.1.155 PP PP e e e e g a g a 00 g a g a 00 tntn tntn ee ee c r e c r e -0-0.1.155 c r e c r e -0-0.1.155 PP PP -0-0.3.3 -0-0.3.3 00 55 1100 1155 2200 00 55 1100 1155 2200 UUnneemmpploloyymmeenntt R Raattee:: T Teexxaass UUnneemmpploloyymmeenntt R Raattee:: T Teexxaass 00.3.3 00.3.3 ss ss tn io tn io 00.1.155 tn io tn io 00.1.155 PP PP e e e e g a g a 00 g a g a 00 tntn tntn ee ee c r e c r e -0-0.1.155 c r e c r e -0-0.1.155 PP PP -0-0.3.3 -0-0.3.3 00 55 1100 1155 2200 00 55 1100 1155 2200 UUnneemmpploloyymmeenntt R Raattee:: N Noorrtthh D Daakkoottaa UUnneemmpploloyymmeenntt R Raattee:: N Noorrtthh D Daakkoottaa 00.3.3 00.3.3 ss ss tn io tn io 00.1.155 tn io tn io 00.1.155 PP PP e e e e g a g a 00 g a g a 00 tntn tntn ee ee c r e c r e -0-0.1.155 c r e c r e -0-0.1.155 PP PP -0-0.3.3 -0-0.3.3 00 55 1100 1155 2200 00 55 1100 1155 2200 QQuuaarrteterrss QQuuaarrteterrss Note: The figure shows the impulse response functions (IRFs) to oil-price shocks in the U.S. state-level panel VAR (Section 4). Solid lines represent median IRFs. Shaded areas represent 16th and 84th percentiles constructed using 5,000 bootstrap replications. Variables are expressed as percent deviation from quadratic trend. B.4