Local Labor Markets Adjustments to Oil Booms and Busts

Transcription

1 Local Labor Markets Adjustments to Oil Booms and Busts Gaetano Basso July 21, 2016 Abstract This paper provides novel dynamic analyses of regional economic performance following global oil price swings. Natural resource-induced labor demand shocks can be positive or negative contrary to other commonly studied shocks that hit in one direction only. Resource-rich regions can respond differently to shocks of similar magnitude but opposite sign, thus affecting long-run growth. Analyzing both timings and pathways towards long-run equilibria I find troubling consequences of oil boom-bust cycles: their incidence tends to be local; busts depress disproportionately the lower end of the income distribution; the reduction in human capital stock during booms may exacerbate the costs of busts. JEL Classification: Q33, R12, R23, J23, J31, J24, E32 Keywords: resource booms, local labor markets, oil shocks, human capital, income distribution Department of Economics, University of California, Davis. Contact: gbasso@ucdavis.edu. Website: I am grateful to Marianne Page and Giovanni Peri for their advice and support from a very early stage, as well as to Oscar Jordà and Doug Miller for invaluable discussions. I also thank Jim Bushnell, Mary Daly, Hilary Hoynes, Elira Kuka, Erich Muelhegger, Michel Serafinelli, Na ama Shenhav, Danny Yagan and seminar participants at UC Davis, University of Essex, NHH Bergen, CERGE-EI, Erasmus University Rotterdam, Banca d Italia, Banque de France, UN Reno, ACLEC at UCLA and UC Berkeley, IZA ESS, Università Cattolica Milan, and frdb Workshop at EIEF Rome for helpful comments on early versions of the paper. All errors are my own. 1

2 1 Introduction After a decade of sustained growth, U.S. oil industry prospects are currently threatened by what has been a dramatic two-year long decline in the price of oil. This may have deep economic consequences for oil-rich regions in the U.S., and it is as yet uncertain to what extent their labor markets will be affected. Who is bearing the incidence of up and down swings in the price of oil? What are the medium to long-run consequences on local residents real wages and employment? This paper analyzes the recent history of U.S. oil-producing local labor markets to address these questions. I study the incidence of oil booms and busts by combining global oil price shocks with spatial variation in resource intensity. The interaction of these two measures jointly determine changes in the local value of natural resources, which in turn trigger local labor demand shifts. Unlike previous studies, I also control for changes in the composition of the population. Furthermore, I estimate the timing of adjustment and the pathways by which adjustment occurs using a novel dynamic analysis that compare short-run responses to oil shocks to longer-run outcomes. 1 There is a large literature and an ongoing debate regarding how countries natural resource endowments affect economic growth (Van der Ploeg, 2011; Frankel, 2012; Aragon, Chuhan-Pole and Land, 2015; Arezki, Ramey and Sheng, 2015; Smith, 2015). More recently, studies have begun to examine the impact of natural resources within countries. In particular, existing research has focused on whether commodity price shocks have productivity spillovers onto local non-extraction sectors (Black, McKinnish and Sanders, 2005a; Allcott and Keniston, 2014; Jacobsen and Parker, 2014). It is currently unclear whether natural resource shocks have symmetric effects across booms and busts, whether they have a similar impact on local labor markets in the short and in the long run, and who is actually bearing the incidence of such shocks. Moreover, it is unclear to date how mechanisms such as investments in physical and human capital, or crowding out of the manufacturing sector (through a local price increase), play out in determining 1 The focus is on the period and is dictated by endogeneity concerns arising in the last fifteen years, in terms of oil price fluctuations, oil reserves and technology adoption. See section 3 for more details. 2

3 the long-run growth of natural resource-rich regions (Black, McKinnish and Sanders, 2005a,b; Michaels, 2010; Allcott and Keniston, 2014). 2 I am able to make headway on answering these unresolved questions by (i) bridging short-run responses and long-run equilibria, (ii) accounting for the effects of local price changes, (iii) looking at the changes in the income distribution, and (iv) analyzing the effects of human capital investment over a full boom and bust cycle. I show that such comprehensive analysis is necessary to understand the incidence of macroeconomic shocks on local economies and to explain the long-run economic fortune of resource-rich U.S. regions. My starting point is the implementation of a dynamic analysis that allows me to describe the evolution of responses to oil price shocks from short-run changes to long-run equilibria. I estimate Cumulated Response (CRs) functions by means of local projections (Jordà, 2005) on annual local labor market employment, earnings per worker and population. This method bridges short and long-run analyses and can provide an additional tool for applied labor economists, who rarely use VAR models (as in Blanchard and Katz, 1992) in the context of serially correlated local labor market shocks. I start by estimating population elasticities since migration decisions can help determine the extent of employment and earnings responses. I find that population net inflows to oil price increases are small, but that there are larger net outflows during a bust, especially in the long run. Local employment increases during booms are also smaller than the declines that follow a bust, and employment responses are larger in the long run than in the short run. In contrast, annual earnings per worker reveal a premium soon after a positive oil shock, but this premium dissipates over the following ten years. Earnings do not fall in the first four to five years after a negative oil shock, revealing some rigidities, but then decline in the longer run. Taking my findings as a whole, labor market outcomes exhibit limited responses in the short run, but these cumulate significantly over time. Local labor markets do respond to global price shocks, although with substantial lags, contrary to previous analyses that 2 In the cross-country context, the crowding out of the manufacturing sector occur through an appreciation of the real exchange rate. This phenomenon is often referred to as Dutch Disease (Frankel, 2012). 3

4 do not bridge short and long run (Allcott and Keniston, 2014). I then turn to an analysis of the long-run incidence of oil shocks. To do this I compare local market-level outcomes across decades, exploiting the monotonic upwards and downwards oil-price trends that occurred between , and , respectively. Workers and capital owners experience similar income gains and losses during booms and busts, but changes in local prices largely undo these gains and losses in real terms. I further analyze the effects of the oil boom and the subsequent bust along the total income distribution. I find that a boom benefits all income percentiles equally, but that a bust has a stronger negative impact below the median, particularly at very low income percentiles. Limited migration together with substantial changes in local nonemployment and an increase in the poverty rate imply that regions rich in oil largely bore the incidence of shocks to the value of their reserves. I find that economic growth in oil-rich labor markets did not catch up with that of similar regions until the late 1990s. I also find that labor markets recover more slowly from a negative shock if sluggish migration is accompanied by higher local nonemployment instead of a decrease in local wages, as is predicted by simple dynamic models of spatial equilibrium (e.g., Blanchard and Katz, 1992). The adverse long-run effects of the 1980 s decline in oil prices were worsened by another troubling consequence of the 1970s oil boom: I show that relative to those living in non-oil-rich labor markets the share of 16 to 22 year olds in oil-rich areas who did not attend any school increased by roughly 4 percent. Previous convincing studies have documented similar contemporaneous phenomena (Black, McKinnish and Sanders, 2005b; Morissette, Chan and Lu, 2015; Cascio and Narayan, 2015; Marchand and Weber, 2015), but unlike existing studies I can also observe how the increase in the fraction of the population with low skill levels spills over onto the local economy in subsequent decades. My results suggest that oil-rich regions have, in recent years, suffered from a lack of skilled labor that would have been complementary to productivity-enhancing technologies that were adopted during the 1980s and 1990s. As further evidence of this I show that oil-rich regions have experienced losses in the number of jobs involving abstract tasks even when the oil price has 4

5 been relatively stable and during a period when the U.S. has experienced a technology boom. This paper also fits into a broader literature on regional adjustments to local labor demand shocks (Moretti, 2011). In particular, despite growing research in recent years (Ganong and Shoag, 2015; Autor, Dorn and Hanson, 2015; Diamond, 2012) the causes of the puzzling and unprecedented decline in U.S. regional income convergence (beginning around 1980) continue to be debated. Oil price boom and bust cycles may explain a part of the lack of convergence that is driven by regions with adverse growth outcomes in the 1980s and 1990s. I show that oil-rich regions have experienced either negative, or limited growth for almost two decades, and that growth in these regions has diverged from growth in local labor markets that in the same decades benefitted from the technological booms. I also speak to conflicting evidence on the role of migration as an equilibrating mechanism in the aftermath of a labor demand shock (Blanchard and Katz, 1992; Yagan, 2014; Monras, 2015; Saks and Wozniak, 2011; Notowidigdo, 2011; Autor, Dorn and Hanson, 2013). By leveraging cleanly identified non-mean-reverting shocks to local labor demand, and allowing for asymmetric responses, I am also able to show that, relative to a positive shock, negative shocks lead to more sluggish migration responses. These results have important implication for policy: the traditional policy recommendation for resource-rich countries is to save a share of the rents acquired from nonrenewable resources during booms, and invest the savings into public infrastructure, education and health services (Hartwick, 1977; Aragon, Chuhan-Pole and Land, 2015). Several states have followed the so-called Hartwick rule by setting up sovereign funds that invest natural resources rents (e.g., Saudi Arabia, UAE, Norway, whose funds finance pensions and national insurance schemes). As I show that negative shocks are borne locally and hinder human capital accumulation at the low end of the distribution, a clear policy goal should be to invest rents at the local level in programs that benefit low-skilled workers. Effective policies might include training programs that target the low-skilled. In addition to increasing local skill levels, such programs would also improve low-skilled workers ability to migrate to other booming regions. 5

6 The remaining sections of the paper are structured as follows. In Section 2, I discuss the existing literature in more details, highlighting what we know, and what we do not know, about oil shocks and local labor markets. Section 3 presents the empirical strategy and discusses my identification assumptions, and concludes by describing the data used in the rest of the paper. Section 4 presents the results. In section 5 I discuss potential pathways leading to long-run effects focusing on the role of human capital investments (section 5.2). Section 6 concludes the paper. 2 What we know (and don t know) about oil shocks and local labor markets Economic theory provides limited insights as to how we should expect natural resource shocks to affect local labor markets. Allcott and Keniston (2014) build a model based on heterogenous firms and sectors which predicts that a resource boom will cause an increase in wages if the labor supply is not fully elastic and an increase in local employment concentrated in the upstream extraction sector, but not in the tradable manufacturing. However, it is an open question whether these shifts will be permanent or temporary, and whether they will be reversed after a bust. Moreover, their model provides no guidance about how labor markets will react to unprecedent large shocks. And, finally, it does not discuss the timing and pathways of adjustments to the long-run equilibrium. Blanchard and Katz (1992) offer a general framework that does allow for an analysis of local labor markets dynamic adjustments to long-run equilibria. They note that equilibria can be restored after negative shocks by out-migration of workers and/or inmigration of firms. However, new firms will only move into the labor market if wages become low enough. The new long-run equilibria thus depend on the short-run relative adjustments of prices and quantities. The best empirical evidence on local labor demand shocks is based on either positive technology shocks (Greenstone, Hornbeck and Moretti, 2010) or negative trade shocks (Autor, Dorn and Hanson, 2013; Dix-Carneiro and Kovak, 2015). Even within these 6

7 literatures the possibility that adjustments to positive and negative shocks might occur along different margins and with different timing has not been addressed. For example, if the boom occurs in areas with substantial labor market slack the adjustment process may be quick and concurrently low nominal wages may induce employment growth along with inflows of firms and capital. After a negative shock, instead, the new long-run equilibria may be reached more slowly if out-migration is sluggish and wage frictions cause local employment conditions to worsen. 3 In the context of natural resources, large global price reversals can determine large structural changes in local economies. Therefore, the setting I explore is convenient to test these questions: oil shocks can be both positive and negative, and they are non-mean-reverting, thus providing a good quasi-experiment to test local labor market adjustments with respect to the existing literature. 4 The empirical literature on the potential asymmetric effects of positive and negative natural resource shocks is not conclusive. Existing studies have not investigated the mechanisms that lead to adverse long-term outcomes, nor have they examined the long-run incidence of natural resource shocks (Black, McKinnish and Sanders, 2005a; Michaels, 2010; Marchand, 2012; Allcott and Keniston, 2014; Jacobsen and Parker, 2014). 5 This paper complements the literature by establishing the relationship between oil prices shocks and local labor markets conditions, and investigating the incidence of these shocks. The papers closest in spirit are Jacobsen and Parker (2014) and Allcott and Keniston (2014). Both of these studies estimate the aggregate consequences of the Recent evidence suggests that mobility is limited following trade shocks (Autor, Dorn and Hanson, 2013), but not after mass layoffs (Foote, Grosz and Stevens, 2015; Huttunen, Moen and Salvanes, 2015). Two recent papers, Yagan (2014) and Monras (2015), suggest that migration is still a prominent mechanisms of local labor market equilibration during the Great Recession. 4 As noted by Kline (2008), the oil price series has a unit root and innovation to the series lead to non-mean-reverting shifts. 5 The literature on within country resource specialization has expanded substantially in the recent period motivated by the technology advancements in the oil and natural gas extraction industry, and by the rise in the price of oil between and However, the most recent literature that leverages the advancements in the hydraulic fracturing technology (or fracking) cannot observe the consequences of an oil bust, and does not investigate the mechanisms that can affect long-run growth, nor how benefits and costs are distributed within labor markets across time (Fetzer, 2014; Feyrer, Mansur and Sacerdote, 2015; Weber, 2012). 7

8 boom-bust period using county-specific panel data. Jacobsen and Parker focus on the impact of oil shocks on income and employment, while Allcott and Keniston focus on productivity spillovers. Taken together these studies suggest that oil booming regions have lower income per capita in the long run than if the boom never occurred, and that while a boom produces productivity gains in the short run there are no permanent effects. Even short term productivity spillovers may not translate into real income gains, however, if local prices adjust. Moreover, the benefits of booms (and the costs of busts) may be shared unequally across the population, or vary with the sources of income. They may also be spread across the country via migration. With respect to these papers, I exploit an innovative methods (CRs by local projections) that bridges short-run responses and long-run equilibria, in the spirit of Blanchard and Katz (1992) s model. Then, once established the main local labor market elasticities to oil price shocks, I exploit the Census microdata to analyze their incidence. Unlike most studies, I also expand the analysis beyond a single region, exploiting the fact that CZs, unlike MSAs, encompass both rural and urban areas of the U.S. 6 I exploit arguably cleaner identification strategy than the one used by some of the existing literature, as my proxy for oil richness does not rely on changes in noisy measures of oil reserves or on oil production, both of which can induce bias in the estimates. Furthermore, my time series variation is determined by factors largely exogenous to U.S. CZs. 7 Finally, this paper is also related to the literature on human capital investments. Changes in human capital accumulation can influence both workers mobility and the long-run success of local economies, helping to explain why the productivity effects found by Allcott and Keniston do not appear to last in the long run. Evidence from the coal 6 Carrington (1996); Black, McKinnish and Sanders (2005a); Michaels (2010) and Jacobsen and Parker (2014) focus exclusively on a single U.S. region: Alaska, Appalachia, U.S. South and Mountain West, respectively. Unlike previous studies, I focus on the entire U.S. CZs that delineate the boundaries of a local economy according to commuting patterns, thus reducing the likelihood of confounding spillovers due to local migration and commuting (Tolbert and Sizer, 1996; Autor and Dorn, 2013). Further details about CZs are provided in the Online Appendix OA2. 7 Results using a shift-share instrument for oil shocks, thus leveraging spatial variation in the 1970 share of oil and mining employment across CZs, indicate elasticities that are on average up to two-three times larger than the baseline (results available upon request). 8

9 boom in rural Appalachia (Black, McKinnish and Sanders, 2005b) and the recent fracking boom in the U.S. and Canada (Morissette, Chan and Lu, 2015; Cascio and Narayan, 2015; Marchand and Weber, 2015) indeed shows increases in high school dropout rates following natural resource shocks. Unlike to these papers, however, I am able to analyze the consequences of these large disinvestments in human capital after a full boom-bust cycle. 8 In particular, I provide suggestive evidence that a fall in human capital investment during the boom may further reduce oil-rich areas technology adoption, and their later attractiveness to firms. 3 Identification of oil-induced labor demand shocks 3.1 Empirical approach The source of identifying variation needed to pin down the effects of oil shocks on local labor market employment, wages and inequality should not be determined by endogenous factors, such as labor supply shifts or local development policies. My identification strategy therefore requires an exogenous source of variation both across localities and over time. I accomplish this by creating a measure that combines cross-sectional variation in oil reserves based on the pre-determined distribution of oil fields, with fluctuations in the price of oil over time, which allow for asymmetric effects during boom and bust periods. Consider the following equation: h ln(y c,t ) = γ t OilRich c h ln(p t ) + φ t + ε c,t (1) where h ln(y c,t ) is the change over the horizon h (between time t and t + h) in the natural logarithm of a outcome y (defined later in the paper in more detail) observed in Commuting Zone c. The difference operator accounts for the characteristics of CZs which are constant over time. φ t, a vector of time dummies, absorbs any shock common to all the CZs in any given time. The main independent variable, OilRich c h ln(p t ), is 8 My results are also consistent with Kumar (2015) who, however, looks only at cross-cohorts schooling outcomes in Texas versus the rest of the U.S. 9

10 an interaction term between the indicator for being an oil abundant CZ and the change in the natural logarithm of the oil price. The differential oil price elasticities in oil-rich versus non-oil CZs are captured by γ t, which I will also allow to vary during boom and bust periods. The identifying assumption in this design is that, absent the oil price increase of the 1970s and the bust of the 1980s and 1990s, CZs that are rich in oil would have followed the same long-run trends as other markets within the same region. In the next subsections I validate my identification strategy by providing detailed information on the spatial variation in oil fields, and by describing the behavior of the oil price over the relevant time period. Baseline summary statistics and graphical pre-trends analysis provide empirical validity that support my assumptions (section 3.4). It is worth emphasizing that by comparing two sets of CZs, oil-rich and non-oil-rich ones I estimate the average effects of the oil shocks between these sets of regions. As emphasized by Allcott and Keniston (2014) and Moretti (2011), such analysis is the policy relevant one to evaluate the effects of natural resource shocks from a locality point of view: the estimated effects will partially include the spillovers from other labor markets, as people and capital moves across regions. However, I provide evidence that the effects of oil shocks I capture are mainly attributable to changes in equilibria of oil-rich CZs and not only due to oil-product demand effects in manufacturing-intensive regions (see the Appendix). In the rest of the paper I will estimate two slightly modified versions of equation (1). First, in section 4.1, I will recover the Cumulated Response functions (CRs) by means of local projections (Jordà, 2005) leveraging yearly fluctuations in the price of oil. I will estimate the following equation cumulating the changes in the outcome variables y at each time horizons t + s, for s = 0,..., H conditional on the observable characteristics at 10

11 time t: ln(y c,t+s ) ln(y c,t 1 ) = s+1 ln(y c,t+s ) = γ s (OilRich c ln(p t )+ + γ s (OilRich c ln(p t )) 1( ln(p t ) < 0))+ + X c,t l δ s,l + α c + ζ r,t + ε c,t+s (2) l=1,2 Differentiating both the left and the right hand side allows me to eliminate CZs fixed characteristics while accounting for spurious correlation issues. In fact, the main local labor market outcomes and the oil price series present unit-roots, as described also in Kline (2008) and Acemoglu, Finkelstein and Notowidigdo (2013) among others. γ s and of (γ s + γ s ), the response to a negative shock, are interpretable as the elasticity of the local labor market outcome y, at each horizon s, to the oil price in oil-rich CZs, with respect to CZs with no oil, conditional on the CZ s economic conditions before the shock hit. The vectors X c,t l (for l = 1, 2) include two lags of the main CZ characteristics (share of manufacturing employment, share of female and rural population) and lag values of the main outcomes (population, employment, annual earnings per worker). 9 Additionally, I control for CZ-specific trends (α c ), and for region-year fixed effects, ζ r,t (the omitted category are CZs with no oil). The errors of the model are serially correlated, both due to the nature of the economic processes, and because of the differencing of the variables. Clustering the standard errors at the state level takes into account both of these factors. The intuition behind a CR function is to observe the cumulative change of the outcome of interest between t and t + s once an exogenous shock hits at time t. 10 Local 9 It is well known since Nickell (1981) that dynamic panel models suffer of endogeneity because the outcome in the previous period, y c,t 1 ( X c,t 1 in equation (2)), is mechanically correlated with the error term. Nickell bias is positively determined by the serial correlation in y and negatively by the length of the series, T : it thus presents serious threats to identification for small T only. The full sample I use spans over 31 years and I have reasons to believe that the bias is negligible. Moreover, I test whether the coefficients change when I exclude the lagged outcome variables from the control set: the results (not reported, but available upon request) reassure about the validity of estimation, suggesting that Nickell bias is not an issue in this context. 10 Analogously, the impulse response, IR, is the change at each t + s for s = 0, 1,... as the shock hits at time t. 11

12 projections have several advantages with respect to the traditional way in which cumulated (and impulse) responses are estimated. The parameters are identified under common identification assumptions used in the applied labor literature such as, the assumption that the oil price shock interacted with the oil-rich indicator is exogenous to local labor market conditions. They also easily accommodate the panel structure of the data; they allow me to investigate non-linearities in the oil price fluctuations; and, finally, both the coefficients and the variance-covariance matrix can be easily estimated with standard regression techniques and packages. 11 The most closely related method to local projections usually used in applied microeconomics to capture this type of responses is an event study where the exogenous event is a one-time occurrence. This latter method, however, does not allow shocks and outcomes to be serially correlated. Once I established the short-run and long-run elasticities to the price of oil by means of local projection, I analyze in more details the long-run incidence of the shocks estimating the following equation on decennial Census data: ln(y r,c,b ) = γ b OilRich r,c ln(p b ) + X r,c,b 1 δ + φ b + ζ r + ε r,c,b (3) The difference operator is now defined over each time period b (b = 1970s boom, 1980s- 1990s bust). The choice of the horizon is dictated by the monotonic changes in the price of oil over these two time periods, as described in section 3.3. The set of control variables, X, includes the beginning of the period workers in the tradable sector, college graduates, females, immigrants, and population residing in rural counties as share of the CZs population, the share of workers occupied in routine intensive occupations (that has been shown to correlate with technology adoption, Autor and Dorn (2013)), as well as the log of the main local labor market outcomes (population, employment, non-employment, wages). The standard error estimates are clustered at the state level in order to account for within state correlations, and the regressions are weighted by the beginning-of-decade 11 Impulse and cumulated responses by local projections have been introduced by Jordà (2005) who provides a discussion of the method, of its consistency and inference properties, as well as a comparison with traditional VAR-based IRs. 12

13 CZ population. 12 In section 4.2 I will establish empirically that the long-run elasticities during the boom and the bust period recovered by equation (3) are analogous to those recovered by the CRs to both positive and negative price changes. 3.2 Oil rich areas in the U.S. The cross-sectional variation in oil richness I exploit comes from the geologically-determined presence of large oil fields across the U.S. Following Michaels (2010), I define large oil fields as those having more than 100 million barrels in oil reserves potentially recoverable and cumulated production (Estimated Ultimate Recovery, EUR) as measured in 1999 (Petzet and Beck, 2000). 13 Crucial to my identification assumption, large oil fields that determine CZs oil richness have been discovered before 1973, i.e., before the first sizeable oil price shock. For two CZs the indicator for oil richness changes in 1976 as new large oil fields were discovered in that year. Excluding these CZs from my sample does not affect the results (see the Appendix). I match this information with a list of all fields and overlaying counties from the U.S. Energy Information Administration (EIA). Thus, I provide a more complete picture of the presence of oil in the U.S. than those in Michaels (2010) and in Acemoglu, Finkelstein and Notowidigdo (2013) by including regions other than just the South, and adding information on small oil fields. I run my main analysis by defining as oil-rich those CZs with at least one large oil field (i.e., a field with more than 100 million barrels in EUR), distinguishing them from those with only smaller oil fields (between 0 and 100 million barrels in EUR) and from those with no oil fields (OilRich c, SomeOil c and NoOil c, respectively). The three sets of CZs are represented in Figure 1: CZs with at least one oil field (in light gray) are spread all over the country, as are CZs with at least one large field (in black), although they are 12 CZs cross state boundaries: I assign CZs to states where the majority of the population resides, as in Autor and Dorn (2013), Autor, Dorn and Hanson (2013) and Cascio and Narayan (2015). 13 Black, Daniel and Sanders (2002) used similarly defined coal reserves in their study of the local effects of coal booms and busts in rural Appalachia. 13

14 slightly more geographically clustered. 14 CZs with small oil fields may experience a mix of effects due to their industry composition. I exclude these CZs from the main analysis, although I include them as an alternative control group in secondary analysis where I test the validity of my identification strategy. Large fields based on EUR threshold and discovered before the 1970s shocks provide the best available approximation to an exogenous source of oil richness variation. This is because: (i) variation in local oil extraction may be endogenous to local economic conditions; (ii) variation in the level of proven or unproven reserves is based on companyreported estimates that are a function of the technology available at the time of appraisal and of the level of the oil price. I limit my sample to the three decades prior to 2000 ( ), at the time I observe my measure of oil richness. Since early 2000s, in fact, the hydraulic fracturing (fracking) activities from the shale reserves in North Dakota, West Texas and Pennsylvania expand significantly. Other states and local governments, however, issued regulatory restrictions, or even banned, the adoption of this technology (e.g., mainly in California and the state of New York). Thus, improvements and endogenous adoption of fracking technology have dramatically and differentially affected fields reserves estimates across states, endogenously changing the oil abundance U.S. regions. Therefore limiting my sample to year 2000 reduces the concerns that my classification of oil-rich U.S. local labor markets is influenced by non-classical error related to technology adoption The results are robust to defining oil richness as a continuous measure based on EUR barrels (see the Online Appendix OA1). The interpretation of the coefficients, however, is less transparent and, unfortunately, EUR for small fields regions is not available. 15 My identification strategy aims at minimizing sources of non-classical measurement error that can bias the results: in this sense it differs from Jacobsen and Parker (2014) and Allcott and Keniston (2014) who leverage variation in oil extraction and in the specific amount of oil reserves, respectively. In particular, Allcott and Keniston (2014) mix different measures of reserves that may capture endogenous reporting and technology adoption by oil companies rather than resource endowments, especially in the latest years (from 2000 on). 14

15 3.3 Oil price fluctuations I combine cross-sectional variation in reserves with time series variation in oil prices as measured by the West Texas Intermediate price (WTI), expressed in real 1999 dollars. 16 The log real price of crude oil between 1950 and 2000 is shown in Figure 2. The price of oil increased substantially in the 1970s, after having been nearly flat for more than two decades. The initial increase between 1973 and the real price increased by around 150 percent in a single year - was driven by an oil embargo proclaimed by the Organization of Petroleum Exporting Countries (OPEC). A second oil crisis occurred in 1979: the Iranian revolution and the explosion of the Iran-Iraq conflict in caused the price to increase abruptly again, reaching its maximum level over the thirty years (around $76 per barrel). The following counter shock was determined by a series of factors, starting with a decline in worldwide aggregate demand that accompanied the late-1970s recessions, the discovery of unexploited reserves in the North Sea, and a dramatic increase in Saudi Arabian production in These factors brought back prices in the range of dollars per barrel by the end of the 1980s. The price of oil declined in the 1990s, with the exception of small rises induced by the Gulf War ( ) and the rise of the Chinese economy in the late 1990s. 17 The major fluctuations in the price of oil during this period largely resulted from disruptions (or upsurges) in the supply of oil. The nature of oil shocks is, however, a highly debated topic in macroeconomics and global demand is known to have played an important role as well (Kilian, 2009; Baumeister and Peersman, 2013). For the purpose of this paper, it is crucial to notice that U.S. regions that specialize in oil extraction are small enough that each of them is a price-taker in the global oil market. 18 Still, 16 All dollar amounts in this section and in the rest of the paper are deflated by the CPI-U All Items index. 17 The events listed in this paragraph correspond to those identified by the EIA as the main drivers of the oil price changes between 1970 and 2000 ( spot_prices.cfm, last accessed on May 31, 2016). 18 Using a structural VAR model Kilian (2009) shows that the nature of oil shocks (e.g., changes in oil supply, oil demand or precautionary/speculative demand) might have different implications in terms of GDP responses at the macro level. Whether local labor markets also respond differently to different sources of price variation is left for future research. 15

16 in the unlikely and extreme case that the increases in the price of oil during the 1970s were entirely due to expansions in aggregate demand among non-oil producing areas of the U.S., my estimates would be biased towards zero. In fact, robustness checks shows that, if anything, my baseline estimates are downward bias when I further control for oil-product demand. 3.4 Data and baseline summary statistics I use two sources of data for my outcome and control variables. In section 4.1, when I analyze the timing of adjustments to oil shocks, I use annual data on county-level population, total employment and annual earnings per worker (deflated by the national CPI) as measured by the Bureau of Economic Analysis Regional Economic Accounts (BEA REA), which I aggregate at the CZ level. 19 In the rest of the paper I use outcome and control variables derived from the 1950, 1970, 1980, 1990 and 2000 population Census public use microdata (Ruggles et al., 2010), which I also aggregate to the CZ level. 20 I mainly focus on outcomes among the working age population (ages 16 64). Population, employment (overall and by sector), and non-employment (i.e., the number of unemployed individuals plus the number of individuals not in the labor force) allow me to analyze labor market adjustments following oil booms and busts, as well as the role of migration and local unemployment in absorbing the shocks. I capture the effects of oil shocks on individual income by estimating impacts on annual earnings, weekly wages (annual earnings divided by the number of weeks worked in the previous year) and income from capital, as well as on total income. These different measures together with the poverty rate allow me to analyze the local incidence of oil shocks. In order to account for changes in the composition of the population, and because the units of observation are CZs, the main wage variable used in the analysis is derived following a common regression-adjusted two-step procedure (as in Notowidigdo, 2011; Beaudry, 19 Unfortunately, the absence of yearly series of local prices and/or housing rents at the CZ-level prevents me from analyzing local real wages at annual frequency. 20 At the time of the analysis the 1960 public use Census microdata did not contain geographic identifiers smaller than a state and it was thus not suitable for my research design. 16

17 Green and Sand, 2014). This wage measure, which I also refer to as the CZ premium, is the estimated CZ fixed effect from a decade-specific individual-level regression in which the log of the weekly wage is the dependent variable and the control variables include both a set of individual level demographic characteristics and a vector of CZ dummies. 21 I use housing rents to capture the responses of local prices to oil shocks. In order to control for potential endogeneity of the local housing market, I measure housing rents by taking the residuals from decade-specific regressions where I regress the monthly net rental costs on building characteristics. I also construct local real weekly wages by subtracting.3 ln(housingrent) from the natural logarithm of weekly wages (Moretti, 2013). This measure assumes that the average share of housing costs in the household budget is 30%, which is consistent with estimates calculated by the BLS (Moretti, 2013; BLS, 2015). Further details on the construction of the variables and on the samples used for both annual and decennial data is are provided in the Online Appendix A2. Table 1 reports the baseline summary statistics as of 1969, before the changes in the price of oil that took place in the 1970s, across CZs with large fields, smaller fields and no oil. In the last column I report the p-values from tests of the difference between the main outcomes means across oil-rich and no oil CZs, conditional on Census division dummies. Oil-abundant CZs tend to have a different industrial mix, being more specialized in the oil and mining sector. They also have a lower share of high school dropouts. Overall, however oil-rich markets are comparable to those with no oil reserves within the same region as of My first differences regression analyses further ensure similarity between CZs because they control for both observed and unobserved CZ characteristics that are constant over time. My identification strategy assumes that oil-rich CZs would have followed the same trends as other CZs, absent any change in the oil price. Two sets of analyses support this assumption. First, in figure 3 we see that oil-rich and non-oil-rich labor markets followed similar trends between 1969 and The pre-trend analysis is validated 21 The results are robust to different specifications of the covariates set that is described in details in the Online Appendix OA2, and to different measures of wage (see the robustness checks in the Appendix. 17

18 also when I test the long-run differences between oil-rich and non-oil-rich CZs between 1950 and 1970: the two sets of CZs do not show differential trends in the main labor market characteristics (see the Appendix). In order to validate my binary definition of oil-richness, I also test whether oil extraction responds to changes in the price of oil differentially across CZs. Figure 4 plots total oil production (log barrels per year) in three states (North Dakota, South Dakota and Wyoming), separately for CZs with large fields and CZs with smaller fields. We can see that production follows the price of oil in both types of CZs, although the extent to which this happens is magnified in oil-rich CZs Main results 4.1 Timing of adjustments to oil shocks: Short and longrun elasticities I begin by estimating Cumulated Response functions (CRs) by means of local projections (Jordà, 2005), which allow me to uncover labor markets dynamic adjustments towards long-run equilibria. Figure 5 reports the estimated population, employment and earnings responses to a 100 log point change in the price of oil along with 95 percent point-wise confidence intervals. 23 In general, the short-run (1 to 2 years) elasticities for all outcomes are smaller than long-run ones, and the adjustments are generally not completed within the first five years after the shock. In the first year after an increase in the price of oil earnings per worker and employment both increase by around 1 percent. Within three years the changes have cumulated to 3 percent. In contrast, the local population does not change significantly in the first two years, which implies that there must be a sizeable decrease 22 I was able to collect publicly available data on county-level oil production, which I aggregate to the CZ, for these three states for a long enough time period. 23 The solid blue lines in the figures represent responses to a positive shock, while the red lines, preceded by a minus sign, represent responses to a negative shock. The confidence intervals are represented by shaded areas. 18

19 in non-employment. Even more striking is that the cumulated differential change in population in oil-rich CZs over a 10 year period, although positive, is extremely limited. Although this result may be surprising at first, it can be rationalized by three factors. First, oil-rich CZs experienced a considerable labor market slack in 1970, which allowed them to easily absorb the positive shock to labor demand (Table 1). This does not imply that there was an absence of internal migration however. Rather, it indicates that the labor demand shock was not large enough to divert migration flows from other regions towards oil-rich CZs. Finally, as documented by Saks and Wozniak (2011), internal migration is pro-cyclical. During the period that oil-rich regions were booming, the rest of the country was entering significant economic slowdown, which may have affected migration towards oil-rich areas. Local labor markets adjust somewhat differently after a negative shock. As is the case for positive shocks, employment declines begin shortly after the shock, but the elasticity is much larger. The long-run elasticity of local employment to a 100 log point change in the price of oil is around 10 percent. Annual earnings per worker do not decline as much in the first four years after a shock, but within five years they have dropped by about 3 percent. Blanchard and Katz s (1992) model suggests that a negative demand shock that is accompanied by only a small reduction in local wages may have lead to extended periods of high unemployment. Indeed, my estimates suggest that population does not begin to respond until two years after the shock. Out-migration only becomes significantly large ten years after the shock. Overall, my estimates indicate that the costs of a negative oil price shocks are borne locally: local labor market encounter frictions in adjusting wages, and out-migration takes place slowly. As a point of comparison, consider the local labor market consequences of another adverse macroeconomic shock: the increase in Chinese imports in the 1990s. Autor, Dorn and Hanson (2013) document that the China syndrome also creates adjustments costs that are borne locally: but while out-migration is very limited in both contexts, wages decline by a larger amount in communities negatively affected by Chinese trade penetration relative to the oil context. It is worth noting, however, that 19

20 Chinese competition does not consist of a series of isolated shocks, but rather a change in trends. My analysis of cumulated responses informs the timing and the size of aggregate labor market adjustments after shocks to the price of oil. However, absent microdata, it is impossible to estimate who benefits and who loses from natural resource shocks. Moreover, changes in the composition of the population might confound the estimated effects on earnings gains and losses. paper. I turn into these issues in the remainder of the 4.2 Oil shocks and local labor markets equilibria I begin my investigation of the long run incidence of oil shocks by first confirming that the elasticities estimated in section 4.1 are comparable with when estimated with decennial Census data. Table 2 reports estimates of γ b from equation (3) for population, employment and weekly wages. All of the reported coefficients are multiplied by 100, and thus represent the responses to a 100 log point change in the price of oil. 24 Given the identification strategy, I expect to observe long-run boom and bust elasticities similar to the ten-year responses estimated by local projections. The unconditional estimates in panel A indicate slightly stronger responses with respect to those in panel B, which include the full set of controls. This result is to be expected since the controls in panel B capture local labor market conditions at the beginning of each period, which have been impacted by previous fluctuation in the price of oil. Overall, the estimates are very similar across panels, and extremely close to the long-run elasticities estimated by the local projections, thus supporting the appropriateness of the long-run identification strategy. This analysis also confirms the striking asymmetry of the employment response: my preferred estimates, in panel B, indicate that the employment change is more than three times larger during a bust than during a boom. While a 100 log point increase in the price of oil leads to a 2 percent increase in employment, a 100 log point decrease leads 24 Across the sample period the oil price increased by 330 percent (in the 1970s), which is equivalent to a 146 log point change. It then decreased by 65 percent between 1980 and 2000, which is equivalent to a 105 log point change. 20

21 to a 9 percent decline. In order to understand the effects of improved labor market opportunities on workers real wages we need to account for simultaneous changes in local prices, which we cannot do in the annual analysis. Column (5) shows that rental housing prices increase by about 5.5 percent after a 100 log point increase in the price of oil. This is a large estimate considering the limited migration response I observe during the boom, and suggests that the supply of housing was imperfectly elastic. Quality-adjusted housing values are not reported in the table, but follow the same pattern as housing rents, thus suggesting the possibility of capital gains for local homeowners. 25 Taking housing prices into account has a striking impact on local real wages: during the boom they increase by only two thirds the amount of nominal wages, but as the price of oil drops, real wages are virtually unchanged because local prices drop as well. Although, taking housing prices into account provides only an approximation of the real wage gains and losses across natural resource cycles, it improves on previous estimates wage analyses, which have not accounted for the endogenous response of local prices. I also break down the two periods indexed by b in equation (3) into three decades now indexed by t (boom corresponding to the 1970s, bust (I) corresponding to the 1980s, and a further milder bust in the 1990s named bust (II). I estimate the following equation: ln(y r,c,t ) = β t OilRich r,c + X r,c,t 1γ + φ t + ζ r + ε r,c,t (4) β t capture the growth in the outcome, y, in each of the three decades (1970s,1980s and 1990s) in CZs that are oil-abundant relative to those that are not. The estimated semi-elasticities β t, one for each decade, allow me to better disentangle along which margin labor markets adjust over time. In particular, it could be that, during a bust, adjustments in the size of the local labor force (through out-migration) take longer than ten years. Panel C shows the results of this analysis. Note that the signs of the coefficients now represent the decennial changes (i.e., semi-elasticities) instead of elasticities to the oil 25 The results are available from the author. 21

22 price. During the 1970s, at the time of the boom, employment increased as much as 7 percent. During the 1980s and 1990s busts employment dropped by a similar amount, but cumulated over the two decades. Non-employment increased during the bust by around 8 percent. The migration response during the bust is concentrated in the second decade ( ) when the working-age population declined by 3 percent. Wages increased 5.4 percent more in oil-rich CZs between 1970 and 1980, whereas the estimated effects during the busts are not distinguishable from zero. The change in housing rents during the boom (8 percent) implies little gain in real wages, which, if anything, slightly declined also in the 1990s. The elasticities over the entire boom-bust cycle and the decennial semi-elasticities of panel C provide a complete picture of the long-run adjustments to oil shocks. Notably, local labor markets are more responsive to negative price changes than to positive ones. The adjustments occur through changes in quantities: out-migration largely reequilibrate the market, but population changes occur also after the first ten years of a bust. 4.3 The incidence of the shocks across the income distribution My finding of limited migration responses has two related, but distinct, implications for the incidence of the oil shocks. First, positive labor demand shocks are not likely to be absorbed exclusively by recent migrants. Second, the impact of negative shocks does not spread into other local labor markets through out-migration, but it is borne locally through higher unemployment. My results also indicate that consistent with (Notowidigdo, 2011) the local population is partially compensated in the aftermath of a negative demand shock by a lower cost of living. It is not clear, however, whether these costs are borne equally by all individuals, or whether some groups experience a bigger loss than others. Table 3 sheds some light on this by looking at total annual earnings, income from capital and poverty rates across time. We see that, during the 1970s boom, the local 22

23 benefits were shared between capital owners and workers. Income from capital, however, dropped by a larger amount during the bust, indicating a somewhat larger elasticity to the price of oil. 26 Not surprisingly, poverty rates behave counter-cyclically, decreasing during busts and increasing during booms. So far, the impact of the oil shocks on the distribution of income is ambiguous across booms and busts. The gains during the boom seem to be shared between workers and capital owners, but higher non-employment and limited migration during the bust suggest that the incidence of the negative shocks may be borne by a limited group of the local population. In order to test empirically the distributional effects, I compare the unconditional quantiles of the log total income distribution, across boom and bust periods. The quantity ln(y) q OilRich c,r,t ln(y) q NoOil c,r,t captures the differential changes over time at each fifth centile of the log total income distribution, q =.5,...,.95, as a function of CZ oil abundance and time period and after accounting for the division (r) average of each centile. 27 The estimated differences, plotted in Figure 6 for each decade, inform the effects of oil shocks along the income distribution. The dashed lines represent the average effects on weekly wages estimated in Table 2. The asymmetry of the results is striking: while the 1970s oil boom shifted total income up at all percentiles by roughly the same amount, the bust hit mostly the low end of the distribution. The results imply that inequality increased more in oil-rich CZs compared to areas of the U.S. with no oil reserves. The 90th-10th log income differential increased by about 10 percent more in oil abundant CZs between 1970 and 2000, with lower-tail inequality (50th/10th) increasing more than upper-tail inequality (90th/50th). The results contribute to our understanding of the distributional impact of oil shocks: specifically, they suggest that the consequences of negative shocks triggered by the fluctuations in the oil price are larger for the lower-end of the distribution than are the benefits deriving 26 Results not reported in the tables also show that both low-skilled and high-skilled workers benefited during a boom. Low-skill workers earnings decreased slightly more during the first decade of the bust, but the overall college premium decreased between 1980 and 2000 (see also section 5.2). 27 Analogously to the other regressions in the paper, the unconditional quantiles are weighted by the beginning of decade CZ population in each bin. 23

24 from a boom. These effects, however, seem to be attributable more to changes in the employment opportunities, rather than changes to educational premia, as investigated in section Explanations for the long-term negative outcomes My results indicate that employment and population responses are somewhat larger during busts than during booms. As in Allcott and Keniston (2014) I find that the effects of positive oil price shocks are not permanent. 28 The deterioration of employment conditions continued up to twenty years after the onset of the 1980s bust, even with more stable oil prices: this suggests that the effects of negative oil shocks may be long lasting. Figure 7 summarizes these findings, showing adjustment dynamics (represented by thick arrows) and equilibria (dots). In the left panel, we see that during a boom the change in the local value of oil shifts labor demand to the right. My results suggest that the supply curve is upward sloping, and not perfectly elastic, as is often assumed in other classical long-run spatial equilibrium models (Rosen, 1979; Roback, 1982). Adjustments during the boom occurred first along the supply curve, leading to an increase in nominal wages. There were also substantive increases in prices that were likely due to stronger demand conditions, increases in construction wages, and inelastic housing supply during the boom. The long-run equilibrium after a positive oil shock corresponds to the point (A) Boom which is associated with higher employment and, to a lesser extent, higher real wages. The right panel summarizes the effects of the oil bust. Starting from the equilibrium (A) Boom the decrease in the price of oil-shifted local labor demand to the left. Employment adjusted initially along the supply curve, leading to higher non-employment, but while nominal wages fell, the decline in prices lead to a smaller decline in real wages. The drop in local prices is compatible with a durable housing 28 On the contrary, changes in Total Factor Productivity (TFP) have permanent effects on local labor markets (Greenstone, Hornbeck and Moretti, 2010; Hornbeck and Moretti, 2015). 24

25 stock, as described in Glaeser and Gyourko (2005) and Notowidigdo (2011). In terms of incidence, income from capital shows an elasticity to the price of oil slightly larger than that of earnings during both booms and busts. Moreover, given sluggish out-migration, the costs of a bust are borne locally. Workers migrate out of oil-rich areas, thus driving up real wages, only twenty years after the original drop in oil prices. 5.1 Dutch disease Why does it take so long for local economies to reach equilibria after a negative oil shock? Can the estimated asymmetries and the incidence of oil shocks be driven by the classical hypotheses discussed in the cross-country natural resource curse literature? The most often discussed among these hypotheses is called the Dutch Disease. 29 According to this hypothesis, positive natural resource shocks crowd out other sectors that drive economic growth (e.g., tradable manufacturing) through a real appreciation of the currency. In the within-country context, this mechanism corresponds to a rise in local prices, which we observe in the U.S. oil-rich CZs during the 1970s boom. Like Allcott and Keniston (2014), the empirical evidence, reported in the Appendix, rejects the Dutch Disease hypothesis because employment in the tradable sector does not decline during the boom, nor it is particularly negatively affected during the bust. Therefore, another mechanism must be responsible for the large negative elasticities of employment and population observed during the bust, and thus for the local economy s inability to sustain long-run growth negatively affecting the lower end of the income distribution. 29 Other hypotheses include institutional deterioration and civil conflicts (Caselli, Morelli and Rohner, 2015) and corruption, especially within newly industrialized countries (Caselli and Michaels, 2013). U.S. have common and stable institutions, and this rules out the institutional channel. Corruption and misallocation of public funds could explain the adverse outcomes observed during the bust, but they would also likely jeopardize the positive gains associated with a boom. My results do not support this hypothesis in the context of the U.S. boom. 25

26 5.2 Human capital investment and high-skill content jobs Changes in the skill composition of the local workforce affect local economies productivity and wages (Moretti, 2004). The 1970s oil shock may have altered the skill composition of the local population during a period when the skill-premium grew substantially (Autor, Levy and Murnane, 2003; Goldin and Katz, 2009). Figure 8 shows the level of schooling by cohort, as a function of the oil price at age 16, as of year 2000 in oil-rich and non-oil-rich CZs. The patterns are striking: the share of individuals with no college education in oil-rich CZs diverges from non-oil CZs right around the time of the oil price increase. Moreover, these do not converge when the oil price drops. Overall, between 1970 and 2000, the share of the population in oil-rich CZs with high school education (or less) grew by around 2 percentage points. In this section, I investigate to what extent changes in human capital investment have contributed to this increase. 30 The existing literature has already documented that the 1970s boom lead to an increase in the number of high school dropouts in coal-rich counties in Appalachia (Black, McKinnish and Sanders, 2005b) and in the oil-rich state of Alberta, Canada, (Emery, Ferrer and Green, 2012); similar phenomena have also been documented in regions that recently experienced the fracking boom in the U.S. (Morissette, Chan and Lu, 2015; Cascio and Narayan, 2015; Marchand and Weber, 2015). These studies, however, have been unable to examine what happens during a bust. The first two columns of Table 4 confirm that during the 1970s booms the share of individuals 16 to 22 years old in oil-rich CZs who were not enrolled in school increased by 2.4 percentage points (around 4 percent) relative to non oil-rich CZs. These estimates are similar when the sample is limited to those who resided in the same CZ five years before the survey (column (2)). Between 1980 and 2000 the share of 16 to 22 years old attending school increased only moderately, suggesting an asymmetric response in human capital investment during booms and busts. In order to gauge the long-term effects of the oil boom-induced drop in human cap- 30 The alternative explanation, i.e., heterogeneous migration across skill groups, has played a minor role. In fact, I do not observe significantly different responses in migration rates across skill groups (see in Table OA1.1 in the Online Appendix). 26

27 ital on local employment and income prospects I also examine the growth of Abstract Task Intensive (ATI) occupations. ATI are occupations whose on-the-job tasks are nonroutine and mainly require problem-solving and creativity rather than manual dexterity or in-person interaction. Managers, professionals, high-skill technicians and creative occupations are the most represented among ATI. 31 The dramatic rise of ATI occupations in the U.S. since 1980 has been well documented (Autor, Levy and Murnane, 2003; Autor and Dorn, 2013). Moreover, it has also been documented that the rise of high-skilled workers in the U.S. lead to local spillovers in both low-skill-intensive and high-skillintensive services (Moretti, 2010; Leonardi, 2015). The rise in high-skilled workers also lead to productivity spillovers (Moretti, 2004). The drop in human capital investment we observe may have therefore had a broader impact on the structure of oil-rich local labor markets. The results in the last two columns of Table 4 show that the share of ATI jobs declined substantially after the oil price dropped relative to other areas of the U.S., while it was stable during the 1970s boom. The coefficients are statistically significant and negative also in the 1990s, although the price of oil was more stable and at lower levels. The coefficients are also robust to including only those who did not migrate in the last five years, and in fact the coefficient for the 1990s decade is larger than the one estimated on the full sample. This result suggests that, while the U.S. was experiencing a technology-driven boom and high levels of growth of ATI employment and earnings, oil-rich areas could not catch up. The contemporaneous contraction in the share of high-skilled workers and the drop of creative jobs likely contributed to a long-lasting economic growth decline of U.S. oil-rich regions. These results, which are new to the literature, are broadly consistent with the argument that a larger presence of low-skill workers reduces the likelihood for firms to adopt technology (Ciccone and Papaioannou, 2009; Beaudry, Doms and Lewis, 2010; Lewis, 2011), thus affecting long-run growth. Moreover, a recent work by Hornbeck and Moretti (2015) indicates that increase in TFP related to technology adoption largely benefit workers (especially the low-skilled) rather 31 For a detailed definition of ATI occupations see the Online Appendix OA2. See also Autor, Levy and Murnane (2003); Acemoglu and Autor (2011); Autor and Dorn (2013). 27

28 than capital owners. The lack of this kind of growth in the 1980s and 1990s may have induced a worsening of the workers employment and income conditions. However, this may also have attenuated the negative effects on inequality described in the distributional analysis. In fact, local economies where technology adoption has been more limited also experienced lower returns to education (Beaudry, Doms and Lewis, 2010). Consistent with this, oil-rich CZs experienced a decrease of the college premium of 1.44 percent when oil price declines by 100 log points (like during the period ; results available from the author). 6 Conclusions Using historical data from the 1970s-1980s oil price shocks, I provide new evidence on the short and long-run effects of shocks to local labor demand. Previous studies have investigated these effects in the context of trade and technology shocks, but existing studies have been constrained by their reliance on events that shift labor demand into a single direction only. This is an important constraint as I show that there are substantive asymmetries in the response to booms and busts. I also contribute to the rapidly expanding, but currently inconclusive literature on the role of natural resources in the local economy. My analyses indicate that in order to determine the impact of such shocks on regional growth and their long-run incidence on local population it is necessary to account for dynamic adjustments, changes in local prices and human capital investment. Estimated Cumulative Response functions by local projections indicate that after a positive shock, employment and annual earnings adjust to their long-run equilibrium within two years. In contrast, it takes much longer to adjust to a negative shock, and the long-run elasticity is 2-3 times larger than the short-run one. Individual gains during the boom were limited by a commensurate increase in local housing rents. Capital earners also gained during the boom, as predicted by a simple static model of local labor markets (Moretti, 2011), and income from capital is more responsive than earnings to price shocks, especially during busts. While there were large 28

29 drops in employment during the 1980s oil bust, real wage declines were modest because of the drop in local prices. Similar to the recent fracking boom human capital investments plummeted during the 1970s, and the large drop in skill acquisition likely contributed in part to the economic decline that oil-rich areas experienced through the late 1990s. As a consequence, workers lacked the necessary skills to take advantage of rising skill premia, possibly increasing their migration and search costs. Income inequality also increased in oil-rich areas relatively to non-oil-rich areas over the period This paper cannot fully evaluate the welfare consequences of oil price shocks, as these can also trigger changes in local amenities and disamenities (e.g., environmental damages, crime) not investigated in this paper. Yet, overall my results suggest that oil shocks can impact the structure of a local labor market up to twenty years after the initial shock. These long lasting impacts may be softened, however, by policies that favor human capital acquisition during the boom, and in its aftermath. Educational policies that aim at reducing high school dropout rates during local booms, and favor skill acquisition during busts, may reduce the long-term adverse effects of negative local demand shocks. Local economies may also be able to combat the negative impacts of resource busts by setting up funds that invest local revenues derived from natural resource booms into activities that support investment in human capital and local infrastructure, thus following a local Hartwick rule (Hartwick, 1977; Van der Ploeg, 2011). Such policies have been rarely implemented in within country contexts and their effectiveness will of course depend on their exact parameters. 32 My results suggest that absent such strategies, U.S. local economies overinvest in the natural oil extraction industry during boom. While this paper focuses only on documenting the impacts of natural resource booms and busts, future research evaluating the effectiveness of such policies is clearly warranted. 32 Two exceptions in the U.S. context are the Alaskan and Texan Permanent Funds. 29

30 References Acemoglu, Daron, Amy Finkelstein, and Matthew Notowidigdo Income and health spending: Evidence from oil price shocks. The Review of Economics and Statistics, 45(4): Acemoglu, Daron, and David H. Autor Skills, Tasks and Technologies: Implications for Employment and Earnings. In Handbook of Labor Economics. Vol. 4, Part B,, ed. O. Ashenfelter and D. Card, Elsevier. Allcott, Hunt, and Daniel Keniston Dutch Disease or Agglomeration? The Local Economic Effects of Natural Resource Booms in Modern America. National Bureau of Economic Research Working Paper Aragon, Fernando M., Punam Chuhan-Pole, and Bryan C. Land The Local Economic Impacts of Resource Abundance: What Have We Learned? The World Bank Policy Research Working Paper NumberWPS7263. Arezki, Rabah, Valerie A. Ramey, and Liugang Sheng News Shocks in Open Economies: Evidence from Giant Oil Discoveries. National Bureau of Economic Research Working Paper Autor, David H., and David Dorn The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market. American Economic Review, 103(5): Autor, David H., David Dorn, and Gordon H. Hanson The China Syndrome: Local Labor Market Effects of Import Competition in the United States. American Economic Review, 103(6): Autor, David H., David Dorn, and Gordon H. Hanson Untangling Trade and Technology: Evidence from Local Labour Markets. The Economic Journal, 125(584): Autor, David H., Frank Levy, and Richard J. Murnane The Skill Content of Recent Technological Change: An Empirical Exploration. The Quarterly Journal of Economics, 118(4): Baumeister, Christiane, and Gert Peersman Time-Varying Effects of Oil Supply Shocks on the US Economy. American Economic Journal: Macroeconomics, 5(4): Beaudry, Paul, David A. Green, and Benjamin M. Sand Spatial equilibrium with unemployment and wage bargaining: Theory and estimation. Journal of Urban Economics, 79:

31 Beaudry, Paul, Mark Doms, and Ethan Lewis Should the Personal Computer Be Considered a Technological Revolution? Evidence from U.S. Metropolitan Areas. Journal of Political Economy, 118(5): Black, Dan A., Terra G. McKinnish, and Seth G. Sanders. 2005a. The Economic Impact Of The Coal Boom And Bust. The Economic Journal, 115(503): Black, Dan A., Terra G. McKinnish, and Seth G. Sanders. 2005b. Tight labor markets and the demand for education: Evidence from the coal boom and bust. Industrial and Labor Relations Review, 59(1): Black, Dan, Kermit Daniel, and Seth Sanders The Impact of Economic Conditions on Participation in Disability Programs: Evidence from the Coal Boom and Bust. American Economic Review, 92(1): Blanchard, Olivier Jean, and Lawrence F. Katz Regional Evolutions. Brookings Papers on Economic Activity, 23(1): BLS Consumer Price Indexes. Bureau of Labor Statistics Handbook of Methods Chapter 17. Carrington, William J The Alaskan Labor Market during the Pipeline Era. Journal of Political Economy, 104(1): Cascio, Elizabeth U., and Ayushi Narayan Who Needs a Fracking Education? The Educational Response to Low-Skill Biased Technological Change. National Bureau of Economic Research Working Paper Caselli, Francesco, and Guy Michaels Do Oil Windfalls Improve Living Standards? Evidence from Brazil. American Economic Journal: Applied Economics, 5(1): Caselli, Francesco, Massimo Morelli, and Dominic Rohner The Geography of Interstate Resource Wars. The Quarterly Journal of Economics, 130(1): Ciccone, Antonio, and Elias Papaioannou Human Capital, the Structure of Production, and Growth. The Review of Economics and Statistics, 91(1): Diamond, Rebecca The determinants and welfare implications of US workers diverging location choices by skill: Harvard University Job Market Paper Updated Dix-Carneiro, Rafael, and Brian K. Kovak Trade Reform and Regional Dynamics: Evidence From 25 Years of Brazilian Matched Employer-Employee Data. National Bureau of Economic Research Working Paper

32 Dorn, David Essays on Inequality, Spatial Interaction, and the Demand for Skills. PhD diss. University of St. Gallen. Emery, Herbert, Ana Ferrer, and David Green Long-Term Consequences of Natural Resource Booms for Human Capital Accumulation. Industrial and Labor Relations Review, 65(3): Fetzer, Thiemo Fracking Growth. Center for Economic Performance CEP Discussion Papers Feyrer, James, Erin T. Mansur, and Bruce Sacerdote Geographic Dispersion of Economic Shocks: Evidence from the Fracking Revolution. National Bureau of Economic Research Working Paper Foote, Andrew, Michel Grosz, and Ann Huff Stevens Locate Your Nearest Exit: Mass Layoffs and Local Labor Market Response. National Bureau of Economic Research Working Paper Frankel, Jeffrey A The Natural Resource Curse: A Survey of Diagnoses and Some Prescriptions. In Commodity Price Volatility and Inclusive Growth in Low- Income Countries., ed. R. Arezki, C. Pattillo, M. Quintyn and M. Zhu. International Monetary Fund. Ganong, Peter, and Daniel Shoag Why Has Regional Income Convergence in the U.S. Declined? HKS Working Paper No. RWP Glaeser, Edward L., and Joseph Gyourko Urban Decline and Durable Housing. Journal of Political Economy, 113(2): Goldin, Claudia D., and Lawrence F. Katz The race between education and technology. Harvard University Press. Greenstone, Michael, Richard Hornbeck, and Enrico Moretti Identifying Agglomeration Spillovers: Evidence from Winners and Losers of Large Plant Openings. Journal of Political Economy, 118(3): Hartwick, John M Intergenerational Equity and the Investing of Rents from Exhaustible Resources. American Economic Review, 67(5): Hornbeck, Richard, and Enrico Moretti Who Benefits From Productivity Growth? The Local and Aggregate Impacts of Local TFP Shocks on Wages, Rents, and Inequality. Unpublished Manuscript. Huttunen, Kristiina, Jarle Moen, and Kjell G. Salvanes Job Loss and Regional Mobility. Institute for the Study of Labor (IZA) IZA Discussion Papers

33 Jacobsen, Grant D., and Dominic P. Parker The Economic Aftermath of Resource Booms: Evidence from Boomtowns in the American West. The Economic Journal. Jordà, Òscar Estimation and Inference of Impulse Responses by Local Projections. American Economic Review, 95(1): Kilian, Lutz Not All Oil Price Shocks Are Alike: Disentangling Demand and Supply Shocks in the Crude Oil Market. American Economic Review, 99(3): Kline, Patrick Understanding Sectoral Labor Market Dynamics: An Equilibrium Analysis of the Oil and Gas Field Services Industry. Cowles Foundation for Research in Economics, Yale University Cowles Foundation Discussion Papers Kumar, Anil Oil Boom Lowers Human Capital Investment in Texas. In Ten- Gallon Economy., ed. Pia A. Orrenius, Jesús Ca nas and Michael Weiss, Palgrave Macmillan US. Leonardi, Marco The Effect of Product Demand on Inequality: Evidence from the United States and the United Kingdom. American Economic Journal: Applied Economics, 7(3): Lewis, Ethan Immigration, Skill Mix, And Capital Skill Complementarity. The Quarterly Journal Of Economics, 126(2): Marchand, Joseph Local labor market impacts of energy boom-bust-boom in Western Canada. Journal of Urban Economics, 71(1): Marchand, Joseph, and Jeremy Weber The Labor Market and School Finance Effects of the Texas Shale Boom on Teacher Quality and Student Achievement. Unpublished Manuscript. Michaels, Guy The Long Term Consequences of Resource-Based Specialisation. The Economic Journal, 121(551): Monras, Joan Economic Shocks and Internal Migration. Sciences Po Discussion paper 1. Moretti, Enrico Workers education, spillovers, and productivity: Evidence from plant-level production functions. American Economic Review, 94(3): Moretti, Enrico Local Multipliers. American Economic Review, 100(2): Moretti, Enrico Local Labor Markets. In Handbook of Labor Economics. Vol. 4, Part B,, ed. O. Ashenfelter and D. Card, Elsevier. 33

34 Moretti, Enrico Real Wage Inequality. American Economic Journal: Applied Economics, 5(1): Morissette, René, Ping Ching Winnie Chan, and Yuqian Lu Wages, Youth Employment, and School Enrollment Recent Evidence from Increases in World Oil Prices. Journal of Human Resources, 50(1): Nickell, Stephen Biases in dynamic models with fixed effects. Econometrica, Notowidigdo, Matthew J The Incidence of Local Labor Demand Shocks. National Bureau of Economic Research Working Paper Petzet, G. Alan, and Robert J. Beck Extended Decline Likely in U.S., Canada Drilling. Oil and Gas Journal DataBook, Roback, Jennifer Wages, Rents, and the Quality of Life. Journal of Political Economy, 90(6): Rosen, Sherwin Wage-based indexes of urban quality of life. In Current Issues in Urban Economics., ed. Peter N. Miezkowski and Mahlon R. Straszheim. Johns Hopkins University Press. Ruggles, Steven, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder, and Matthew Sobek Integrated Public Use Microdata Series: Version 5.0. Minneapolis: University of Minnesota Machine-readable database. Saks, Raven E., and Abigail Wozniak Labor Reallocation over the Business Cycle: New Evidence from Internal Migration. Journal of Labor Economics, 29(4): Smith, Brock The resource curse exorcised: Evidence from a panel of countries. Journal of Development Economics, 116: Tolbert, Charles M., and Molly Sizer US Commuting Zones and Labor Market Areas: A 1990 Update. USDA Economic Research Service Staff Paper Van der Ploeg, Frederick Natural Resources: Curse or Blessing? Journal of Economic Literature, 49(2): Weber, Jeremy The Effects of a Natural Gas Boom on Employment and Income in Colorado, Texas, and Wyoming. Energy Economics, 34: Yagan, Danny Moving to Opportunity? Migratory Insurance over the Great Recession. Harvard University Job Market Paper. 34

35 Figures Figure 1: Oil Abundant U.S. Local Labor Markets Note: The darker CZs have at least one large oil field (EUR 100M bbl.) as of 1999; light grey have only smaller fields; white CZs have no oil. Source: Petzet and Beck (2000) Figure 2: Real Oil Price, Pre-period Boom Bust (I) Bust (II) 4 log(oilprice) Note: The figure plots the annual average oil price (log) (real 1999 $ per bbl, WTI ). Source: FRED,

36 Figure 3: Population and employment trends in U.S. CZs 11 Avg Jobs (log) Avg Population (log) Oil Rich No oil CZ Some oil CZ log(oil Price) (R Ax) Oil Rich No oil CZ Some oil CZ log(oil Price) (R Ax) Note: The figure plots the average log employment (left) and log population (right) in oil-rich CZs (dark line), CZs with small fields (grey line), and CZs with no oil fields (light grey line). The top graphs plot the time series for the entire sample periods, while the bottom graphs zoom on the pre-period Oil rich CZs have at least one large oil field (EUR 100M bbl.) as of The light grey dashed line represents the log oil price (annual average, real 1999 $ per bbl, WTI ). Source: BEA REA, , and FRED,

37 Figure 4: Oil production and the price of oil, by oil abundance of the CZs (ND, SD, WY) 19 Total Oil Production (log bbl/year) Oil Rich CZs Some Oil CZs log Oil Price (R ax) Note: The figure plots the total production in oil-rich CZs (dark) and other CZs with no oil (grey) in North Dakota, South Dakota and Wyoming, as well as the price of crude oil (light grey dashed line) for the period Oil rich CZs have at least one large oil field (EUR 100M bbl.) as of The oil price is the log annual average (real 1999 $ per bbl, WTI ). Source: ND, SD, WY agencies websites and FRED,

38 Figure 5: Cumulated Responses of main local labor markets outcomes to positive and negative shocks Panel A. Population (log) Panel B. Total employment (log) Panel C. Annual earnings per worker (log) Note: The figure plots separately for each outcome the Cumulated Response functions (CRs) over a 10 years horizon estimated by local projections (see equation (2)). The dark line plots the CR to a positive oil price shock, γ, while they grey line plots the CR to a negative price shock,-(γ + γ). The shaded areas represent 95 percent pointwise confidence intervals robust to state clustering. 38

39 Figure 6: The effect of the oil boom and bust across the total income distribution Total income (pct) D D D Note: The figure plots the differences between oil rich and non-oil log total income distribution at each 5th-centile separately for each decade, accounting for common division and time effects in each bin. The differences are weighted by the population in each CZs-specific bin. The dashed lines represent the average effects on weekly wages as reported in Table 2. 39

40 Figure 7: Local labor markets equilibria: boom (left) and bust (right) Note: The figure plots the employment and nominal wages long-run equilibria (dots) and the dynamic adjustments (thick arrows) in oil-rich local labor markets after an oil boom (left panel) and after an oil bust (right panel). Discussion of this framework is presented in sections 2 and

41 Figure 8: Percent of population with at most 12 years of schooling (price of oil at age 16) % Less than college Year of birth + 16 Oil Rich CZs No Oil CZs Oil Price (L ax) 2.5 Note: The figure plots the percentage of population with at most 12 years of schooling living in oil-rich CZs (dark) and other CZs with no oil (grey) as measured in the 2000 Census, by cohort when each cohort was 16 years old. The graph indicates that, by year 2000, cohorts living in oil-rich CZs and exposed to a higher price of oil around age 16 have a larger share of people without college education than the same cohorts living in non-oil-rich CZs. Oil rich CZs have at least one large oil field (EUR 100M bbl.) as of The oil price is the log annual average (real 1999 $ per bbl, WTI ). Source: Census, 2000, and FRED,

42 Tables Table 1: CZ Summary statistics by oil abundance, 1970 Oil Rich CZ Smaller Fields CZ Other CZ p-val. (1)=(3) (1) (2) (3) (within division) Census data Population , , , (521,416) (439,693) (356,819) High School or less (% Pop ) (5.6) (5.5) (5.4) Employed (% Pop ) (3.3) (3.5) (3.4) Oil&Mining (% Total Empl) (2.8) (2.1) (1.2) Tradable Empl (% Total Empl) (5.3) (9.4) (9.4) Construction,Transportation, Utilities (% Total Empl) (1.9) (2.3) (1.9) Local Services (% Total Empl) (4.3) (7.8) (8.2) Weekly Wage (108.0) (99.3) (109.1) Housing Rent (123.1) (105.8) (115.3) Poverty Rate (6.2) (5.1) (5.4) % In School (16-22) (3.9) (4.2) (5.1) % Abstract Task Intensive Jobs (2.7) (3.2) (3.9) Oil data (EIA, Oil&Gas Journal) Oil EUR (M bbl) 1,670,221 (2,796,844) Number of Oil Fields (84) (36) 1st Large Field: CZ Median [min,max] 1926 [1880,1976] 1st Field (any): CZ Median [min,max] [1861,1970] [1859,1990] CZs Rural CZs Note: The table reports the means and the standard deviations (weighted by the population) of the CZ outcomes analyzed in the paper, and the main characteristics of the oil fields (in square brackets the first and the last. The data, described in the Online Appendix OA2, come from the 1970 Public Use Microdata Census (Ruggles et al., 2010) aggregated at the CZ level, the list of large oil fields published by Petzet and Beck (2000) and the EIA County Master List. The CZ is defined as oil-rich if it has at least one large oil field (> 100M bbl of EUR), and with some oil if it contains only smaller oil fields (< 100M bbl of EUR). The p-values in the third column are for a test of the equality of the means between oil-rich CZs and other CZs with no fields controlling for Census division fixed effects. I define as rural the population of the counties in each CZ that do not belong to a MSA as of

43 Table 2: The impact of oil shocks on local labor markets outcomes Pop (16-64) Empl Not Empl Wkl Wages House Rents LocReal Wages (1) (2) (3) (4) (5) (6) A: 100 ln(y), unconditional γ (eq.(3) ) Oil Rich* ln(p (70s) ) (3.26) (3.35) (3.68) (2.66) (3.25) (3.47) Oil Rich* ln(p (80s 90s) ) (3.27) (3.68) (4.89) (3.17) (5.02) (4.00) γ 70s = γ 80s 90s (p-val.) Mean Y R Observations B: 100 ln(y), conditional γ (eq.(3) ) Oil Rich* ln(p (70s) ) (2.57) (2.62) (3.04) (1.85) (2.31) (1.95) Oil Rich* ln(p (80s 90s) ) (4.23) (4.40) (4.46) (1.76) (3.29) (2.18) γ 70s = γ 80s 90s (p-val.) Mean Y R Observations C: 100 ln(y), conditional β (eq.(4) ) Oil Rich (1 = Boom (70s) ) (3.05) (3.23) (3.31) (3.09) (4.29) (2.70) Oil Rich (1 = Bust(I ) (80s) ) (3.07) (3.86) (2.16) (3.59) (12.14) (1.50) Oil Rich (1 = Bust(II ) (90s) ) (2.65) (2.95) (3.08) (1.42) (7.31) (1.32) β 70s = β 80s (p-val.) Mean Y R Observations Note: Standard errors (in parentheses) are clustered at the state level. All regressions are weighted by the beginning of decade CZ population. Outcomes variables are in expressed in logarithms. Each column in panel A and B reports the estimated elasticities (γs) from equation (3), while in panel C the estimated semi-elasticities (βs) from equation (4). The means at the bottom of the tables are for the level of the outcomes. The population is the count of active (i.e., non attending school) inhabitants of each CZ aged The weekly wage measure reported in the table is the CZ premium described in the text. The controls in Panel B and C include the beginning of decade employment in the tradable sector, college graduates, females, immigrants and rural population as shares of the CZ population, and the share of employment occupied in routine intensive occupations, as well as the lag values of the main CZs characteristics (working age population, employment, average wages, and average housing rents). All regressions include Census divisions dummies and time fixed effects. The indicator of oil abundance (OilRich) equals one for those CZs with at least one oil field with at least 100M bbl of EUR as of The excluded category is CZs with no oil fields. Further details in the text. ***, **, * indicate significance at 1-percent, 5-percent and 10-percent level, respectively. 43

44 Table 3: The impact of oil shocks on total income, income from capital and poverty. Annual Earnings Capital Inc Poverty Rate (1) (2) (3) Oil Rich (1 = Boom (70s) ) (3.35) (2.75) (0.84) Oil Rich (1 = Bust(I ) (80s) ) (3.16) (2.03) (0.99) Oil Rich (1 = Bust(II ) (90s) ) (1.35) (3.45) (0.66) β 70s = β 80s (p-val.) Mean Y R Observations Note: Standard errors (in parentheses) are clustered at the state level. All regressions are weighted by the beginning of decade CZ population. Each column reports the estimated semi-elasticities (βs) from equation (4) calculated within education groups. The means at the bottom of the tables are for the level of the outcomes. The population is the count of active (i.e., non attending school) inhabitants of each CZ aged 16-64, by education level. The controls include the beginning of decade employment in the tradable sector, college graduates, females, immigrants and rural population as shares of the CZ population, and the share of employment occupied in routine intensive occupations, as well as the lag values of the main CZs characteristics (working age population, employment, average wages, and average housing rents). The income from rent and dividend variable in the Census microdata is not consistently defined throughout the sample period: hence, I use the decennial CZ difference from the BEA REA. All regressions include Census divisions dummies and time fixed effects. The indicator of oil abundance (OilRich) equals one for those CZs with at least one oil field with at least 100M bbl of EUR as of The excluded category is CZs with no oil fields. Further details in the text. ***, **, * indicate significance at 1-percent, 5-percent and 10-percent level respectively. 44

45 Table 4: Human capital investment and abstract-task jobs during booms and busts In School (% Pop) ATI Jobs (% Empl) (1) (2) (4) (5) No Mig All No Migrants Oil Rich (1 = Boom (70s) ) (0.77) (0.76) (0.42) (0.42) Oil Rich (1 = Bust(I ) (80s) ) (1.82) (1.97) (0.33) (0.35) Oil Rich (1 = Bust(II ) (90s) (1.18) (1.14) (0.31) (0.39) β 70s = β 80s (p-val.) Mean Y R Observations Note: Standard errors (in parentheses) are clustered at the state level. All regressions are weighted by the beginning of decade CZ population. Each column reports the βs from equation (4). The means at the bottom of the tables are for the level of the outcomes. The controls include the beginning of decade employment in the tradable sector, college graduates, females, immigrants and rural population as shares of the CZ population, and the share of employment occupied in routine intensive occupations, as well as the lag values of the main CZs characteristics (working age population, employment, average wages, and average housing rents). All regressions include Census divisions dummies and time fixed effects. The indicator of oil abundance (OilRich) equals one for those CZs with at least one oil field with at least 100M bbl of EUR as of The excluded category is CZs with no oil fields. Further details in the text. ***, **, * indicate significance at 1-percent, 5-percent and 10-percent level respectively. 45

46 Appendix In this section I include additional results that prove the robustness of my findings. Dutch disease and other mechanisms that may lead to longterm busts Here I test how the sectors of the local economy respond to the oil booms and busts. This exercise allows me to rule out the presence of a Dutch Disease, that is the pull of resources from tradable sectors into the natural extraction sectors due to a local price increase, in the U.S. during the period I define four main sectors: oil and mining, tradable (which consists of manufacturing and agriculture), construction, transportation and utilities, and other local services. These are mutually exclusive and sum up to the total CZ employment. I define the tradable sector as manufacturing plus agriculture, following?. 33 In Table A1, panel A, I find that the extraction sector experienced large responses. Following Black, McKinnish and Sanders (2005a), I then calculate the employment multipliers for each sector j using a 2SLS estimator. I instrument the endogenous number of oil and mining jobs, weighted by the share of oil-to-sector j employment, with the oil abundance indicator OilRich c interacted with time dummies. I perform this analysis only on the boom decade (1970s) and the 1980s bust decade (stacked), because of little jobs destruction in the oil and mining industry in the 1990s. The first-stage F-statistics vary between 5.6 and 6.3 depending on the sectors analyzed. These effects most likely represent an upper bound because of weak first-stages. In other specifications (available upon request) where I do not stack the two decades the jobs multipliers are slightly smaller, but in line with the baseline results and the first-stages are much more robust (F-stats are always greater than 10). The estimated coefficients, which can be directly interpreted as the number of job created (lost) in each sector for each job created (destroyed) in the oil and mining sector during the boom (bust), are similar to those estimated by Allcott and Keniston (2014) and Black, McKinnish and Sanders (2005a). The largest multipliers from the oil industry occurs in the construction, transportation and utility sectors, which provide intermediate local inputs to the mining sector. I find evidence of smaller (but positive) effects in the tradable sector, as well as in other local services. The effects during the bust are slightly larger in magnitude, but not statistically different from the boom s multipliers. 34 As discussed also in the main text, the empirical evidence thus rejects the Dutch Disease hypothesis, as the tradable sector does not lose jobs during the boom, nor it is particularly negatively affected during the bust. 33 For further details about variable definition see the Online Appendix OA2. 34 The F-stats vary between 5.6 and 6.3 depending on the sectors analyzed: thus, these effects most likely represent an upper bound because of weak first-stages. In other specifications (available upon request) where I do not stack the two decades the jobs multipliers are slightly smaller, but in line with the baseline results and the first-stages are much more robust (F-stats are always greater than 10). 46

47 Another leading hypothesis in the cross-country natural resource curse literature is that positive natural resource shocks can induce institutional deterioration (and even conflicts, Caselli, Morelli and Rohner, 2015), or corruption at the local level (within a newly industrialized country like Brazil, Caselli and Michaels, 2013). Although these effects have negative impacts on local economic performance, U.S. states and counties have common and stable institutions, and this would lead us to rule out the institutional channel. Local corruption and misallocation of public funds could explain the adverse outcomes observed during the bust, but they would also likely jeopardize the positive gains associated with a boom, as has been observed in the case of off-shore oil discovery in Brazil (Caselli and Michaels, 2013). My results do not support this hypothesis in the context of the U.S. boom. Specification and robustness checks In this section I present results that confirm the robustness of my findings. In Table A2 I report the results from various checks, along with the baseline results of the long-run specification with the full set of controls (panel A). In panel B, I change the control group now including only CZs with small oil fields only. Using small oil field regions has an advantage in that these may resemble oil-rich regions more closely, thus serving as a better control group. However, due to the presence of oil in these regions we expect the estimates to be attenuated. The point estimates and the signs of the coefficients in panel B are largely similar to those of panel A, with the exception of wages and local real wages, which are now not statistically different from zero. Oil-abundant CZs are thus unique in their dependence on the price of oil, experiencing negative long-run consequences from oil price fluctuations. One might also question whether the estimates in Tables 2 and 3 are driven by the effects of the oil shock in oil-abundant areas or whether they are driven by other changes occurring in control CZs. For example, an increase in the price of oil could depress nonoil-rich CZs by depressing the industrial and residential demand for oil and oil-related products (e.g., automotive), or it could induce sectoral shifts that affect treatment and control labor markets differentially. 35 In order to better capture these demand-side confounders I estimate the effects of oil demand while also controlling for the beginning of decade average CZ household expenditures in gasoline and other fuels, and the share of employment in motor vehicles and other transportation industries. I allow the effects to differ in each decade by interacting the variables with decade dummies. The results in panel C are again qualitatively unchanged, apart from the coefficients on the housing rents that are now larger in magnitude. A related source of concern could be that monetary policy and exchange rate fluctuations due to fluctuations in the price of oil affect local labor markets differently depending on their share of export-oriented industries, although the average effects on the economy 35? find that oil shocks cause considerable intra-sector reallocation, and larger job destruction in capital and energy intensive industries. See? for a recent review of the oil shocks and the effects on industrial reallocation. 47

48 are already captured by the decade fixed effects. Low (high) interest rates have been historically associated with low (high) oil prices (?), which could pose a threat to my identification strategy. In the main specification I already control for the lag values of the tradable sectors employment share in order to capture the exposure of a CZs to the aggregate demand conditions at the beginning of the decade. I also estimate a more flexible specification, in which the controls (including the employment share in the tradable sector) are allowed to differ in each decade (by interacting them with time dummies). The results, reported in panel D, are consistent with those in the baseline specification. Moreover, the trends in the dollar s exchange value do not align with decennial trends in the oil price: the Federal Reserve Trade-Weighted U.S. Dollar Index decreased by 5 percent in the 1970s and by 14 percent in the 1980s, but increased by 15 percent in the 1990s. This reduces the concerns about this potential threat. 36 Similarly, a concern might be that instead of the consequences of oil price swings, I am capturing state-decade or region-decade-specific confounders. In panel E, I control for state-by-decade fixed effects now including all the 722 CZs. The excluded set of CZs is still those with no oil fields, but I now also control for time-specific indicators for CZs with small oil fields. This specification captures state-specific shocks in each decade, including the adoption of polices that might be related to local labor markets outcomes or oil extraction regulations. The estimates are largely consistent with the baseline results, although precision is largely lost and the population response during a boom shows a much larger point estimate. In panel F, I include Census division-by-decade fixed effects to control for contemporaneous shocks within each of the nine U.S. divisions, excluding again the CZs with small fields. In this specification, the effects on wages and local prices during booms and busts are imprecisely estimated, while the other outcomes (population, employment and non-employment) present patterns similar to the baseline estimates. It is worth noting that the inclusion of these additional controls increases the magnitude of the 1990s estimates, confirming that oil-rich regions experienced negative long-run consequences from the oil bust in terms of lower population, employment, and a larger drop in wages and housing rents. I perform one additional test to make sure that my results are not picking up mainly effects on the control group. I estimate the local projection main specification on the full sets of CZs, including an interaction of the oil price with each of the three indicators of oil richness, OilRich c, SomeOil c and NoOil c, thus omitting the year fixed effects. In this specification the three interaction terms will also capture annual changes in the outcomes that occur for reasons other than the changes in the oil price, and thus they should be interpreted with some caution. Yet, they give me a sense of how much the total effect can be attributed to the oil-rich CZs, and of how much of the total effect results from changes in the labor market conditions of other CZs. Figure A1 shows that long-run population and employment elasticities can be largely attributed to changes in oil-rich CZs. The effects on earnings per worker are more nuanced since CZs with small oil fields and those with non-oil suffer large losses after a positive oil price shock, which 36 The Trade-Weighted U.S. Dollar Index is the TWEXM series retrieved from FRED, Federal Reserve Bank of St. Louis (last accessed on July 20, 2015). 48

49 is explained by the contemporaneous realization of the oil boom in oil-rich areas of the U.S. and of downturns in other regions. My estimates are robust to the choice of control group and confirmed by my specification checks. However, I also perform two back-of-the-envelope calculations that can help provide bounds on the extent to which my main estimates are driven by the negative effects of oil shocks in CZs with no oil. To date macroeconomists do not agree on how measuring oil shocks at an aggregate level, nor which are the effects of such shocks on aggregate outcomes (??). The first calculation is based on macroeconomists most recent estimates, which suggest that GDP growth responses to an increase in the price of oil are at most 2-3 percent for the late 1970s oil price spikes (over a two years horizon. See? and?). I estimate that during the 1970s oil boom, total income increased by an additional 6 percent in oil-rich CZs. A lower bound for this estimate would then be around 2.6 percent, assuming that GDP dropped by 3 percent over three years in the late 1970s, and that the whole GDP drop is due exclusively to the 332 CZs with no oil reserves in the main sample (and it has not been captured by the time effects). Similarly, I provide a lower bound for the oil-rich CZs employment based on a simulation of state-level employment responses to oil price changes provided by?. Brown and Yücel estimate that employment in oil-rich states increases by a.46 percent for a 10 percent increase in the price of oil. My baseline estimates suggests an employment elasticity of around 0.25 percent to an analogous shock. Based on this comparison, my baseline calculation would underestimate the effects of an oil shock on the employment of oil-rich CZs. Finally, in panel G, I report estimated coefficients from a model that is estimated over the pre-period ( ), when the price of oil was stable. 37 Based on the identification assumption we expect to observe coefficients that are close to zero and not statistically significant, since oil price changes were absent in this period. Indeed, the estimates are smaller in magnitude and more imprecisely estimated than all the other main estimates, further confirming the validity of my identification strategy. The Online Appendix OA1 report other robustness checks where I construct wage measures other than the one reported in the main specification, I limit my sample to rural areas rich in oil (thus comparing my results to those of Black, McKinnish and Sanders (2005a) and Jacobsen and Parker (2014)), and I test other specifications and definitions of the sample. Online Appendix OA1 also reports the checks and falsification exercises on the dynamic analysis. All the checks qualitatively confirm the main results reported in the paper. 37 Housing rents were not been collected in the 1950 Census. 49

50 Table A1. Oil and mining sector employment and local multipliers A: Oil and mining employment, eq.(4) 100 log % of TotEmpl (1) (2) (3) Oil Rich Boom (1970s) (24.09) (19.51) (0.42) Oil Rich Bust (1980s) (9.31) (9.52) (0.24) Oil Rich Bust(II ) (1990s) (10.69) (7.20) (0.17) β = β (p-val.) Controls No Yes Yes Mean Y R B: Employment multipliers (2SLS, see notes) Tradable Cnstr&Util Oth Services (1) (2) (3) Oil Rich, multiplier Boom (70s) (0.65) (0.41) (1.10) Oil Rich, multiplier Bust (80s) (0.82) (0.87) (2.46) mult boom = mult bust (p-val.) Note: Standard errors (in parentheses) are clustered at the state level. All regressions are weighted by the beginning of decade CZ population. In panel A, each column reports the βs from equation (4) on oil mining employment in logs (columns (1) and (2)) and as a share of local employment. The set of controls in column (2) and (3) includes the beginning of decade employment in the tradable sector, college graduates, females, immigrants as shares of the CZ population, the share of CZ rural population in 1970 interacted with time dummies and the share of employment occupied in routine intensive occupations (Autor and Dorn, 2013). In panel B, I estimate separately for each sector a 2SLS model where oil and mining employment, weighted by the oil sector-to-sector j share at the beginning of each decade, is instrumented by an interaction of oil richness and time indicators. All regressions include Census divisions dummies and time fixed effects. The indicator of oil abundance (OilRich) equals one for those CZs with at least one oil field with at least 100M bbl of EUR as of The excluded category is CZs with no oil fields. See the Online Appendix OA1 for more details. ***, **, * indicate significance at 1-percent, 5-percent and 10-percent level respectively. 50

51 Table A2. Robustness checks: Main specification Pop (16-64) Empl NonEmpl Wkl Wages House Rents LocReal Wage (1) (2) (3) (4) (5) (6) A: Main specification Oil Rich Boom (1970s) (3.05) (3.23) (3.31) (3.09) (4.29) (2.70) Oil Rich Bust (1980s) (3.07) (3.86) (2.16) (3.59) (12.14) (1.50) Oil Rich Bust(II ) (1990s) (2.65) (2.95) (3.08) (1.42) (7.31) (1.32) β = β (p-val.) R Observations B: Small oil fields CZ as control group Oil Rich Boom (1970s) (2.16) (2.22) (2.55) (3.08) (4.18) (2.35) Oil Rich Bust (1980s) (3.49) (4.19) (2.03) (2.80) (9.40) (1.44) Oil Rich Bust(II ) (1990s) (2.04) (2.02) (2.52) (2.67) (7.70) (0.97) β = β (p-val.) R Observations C: Controlling for oil demand Oil Rich Boom (1970s) (7.85) (8.30) (7.09) (2.57) (5.25) (2.60) Oil Rich Bust (1980s) (3.16) (3.94) (1.97) (2.96) (11.65) (1.77) Oil Rich Bust(II ) (1990s) (3.30) (3.50) (3.80) (1.24) (5.56) (1.00) β = β (p-val.) R Observations D: Rural CZ only Oil Rich Boom (1970s) (5.73) (6.10) (5.87) (1.78) (5.39) (1.91) Oil Rich Bust (1980s) (5.53) (6.14) (5.04) (1.95) (6.62) (1.95) Oil Rich Bust(II ) (1990s) (4.59) (4.59) (4.93) (1.06) (5.54) (1.35) β = β (p-val.) R Observations

52 Table A2 (cont.): Robustness checks: Main specification Pop (16-64) Empl NonEmpl Wkl Wages House Rents LocReal Wage (1) (2) (3) (4) (5) (6) E: All CZs and state-by-decade FE Oil Rich Boom (1970s) (5.83) (6.06) (5.65) (2.00) (5.44) (1.65) Oil Rich Bust (1980s) (2.97) (2.93) (3.19) (2.23) (8.01) (1.25) Oil Rich Bust(II ) (1990s) (2.24) (2.24) (2.79) (1.66) (4.29) (1.01) R Observations F: Division-by-decade FE Oil Rich Boom (1970s) (6.08) (6.40) (6.06) (2.25) (5.02) (2.34) Oil Rich Bust (1980s) (2.99) (3.15) (2.87) (2.82) (8.58) (1.88) Oil Rich Bust(II ) (1990s) (1.95) (1.91) (2.65) (1.19) (4.63) (1.21) β = β (p-val.) R Observations G: Pre-period ( ) Oil Rich P re period (50 60s) (5.76) (5.75) (6.10) (1.76) R Observations Note: Standard errors (in parentheses) are clustered at the state level. All regressions are weighted by the beginning of decade CZ population. Each column reports the estimated semi-elasticities (βs) from equation (4). The controls include the beginning of decade employment in the tradable sector, college graduates, females, immigrants and rural population as shares of the CZ population, and the share of employment occupied in routine intensive occupations, as well as the lag values of the main CZs characteristics (working age population, employment, average wages, and average housing rents). All regressions include Census divisions dummies and time fixed effects. The indicator of oil abundance (OilRich) equals one for those CZs with at least one oil field with at least 100M bbl of EUR as of The excluded category is CZs with no oil fields. ***, **, * indicate significance at 1-percent, 5-percent and 10-percent level respectively. 52

53 Figure A1. CRs of local labor markets outcomes, all CZs Panel A. Population (log) Panel B. Total employment (log) Panel C. Annual earnings per worker (log) Note: The figure plots separately for each outcome the Cumulated Response functions (CRs) over a 5 years horizon estimated by local projections. Differently from equation (2) and from Figure 5, I now estimate the CRs on the entire set of CZs, thus omitting year FE. The dark line plots the CR to an oil price shock for large oil field (EUR 100M bbl.) CZs; the grey line for CZs with only smaller fields; the light gray line for CZs with no oil. The shaded areas represent 95 percent pointwise confidence intervals robust to state clustering. 53

54 For Online Publication Online Appendix OA1: Additional Results Checks and falsification exercises on the dynamic analysis The richness of the time series allows me to perform two further tests with respect to the CRs estimates. First, with respect to equation (2), I include a spline in the log oil price change (Figure OA1.1). The idea is to examine whether extreme positive or extreme negative shocks have different effects on local economies with respect to normal positive and negative price fluctuations. The results indicate that normal price changes, i.e., between the 33rd and the 66th percentile of the log price distribution, drive most of the labor market responses. This result is consistent with rational agents who perceive isolated and large increases and decreases in the price of oil as temporary thus not responding to extreme shocks. I also perform a falsification test replacing the annual changes in the price of oil with the price of grain. For this exercise I take the residual of a regression of the log of the grain price on the log of oil prices to account for the correlation between the two series. Then, I estimate equation (2) substituting log(p t ) with log( GrainP ˆ t ), whose coefficient captures the elasticity to a grain price shock. 38 If we observe a systematic relationship between a grain price shock and local labor market outcomes in oil-rich CZs, then I must be capturing shocks other than those triggered directly by the price of oil. 39 The results, which are reported in Figure OA1.2, indicate that the responses of oil-rich CZs were not systematically different from those of other CZs with no oil, both in the short and in the long run. The only exception is that negative grain shocks induced an increase in annual earnings per worker. This result is odd, but it does not necessarily invalidate my identification strategy since it is the opposite in direction from what one would worry about. Further specification checks, such as using CZ-clustered standard errors, excluding region-by-year fixed effects and CZ-trends, or using different definitions of the log oil price, are consistent with the baseline results (available upon request). 38 The grain price series is the World Bank Grains index that includes barley, maize, rice and wheat and it has been converted to 1999 real dollars (IGRAIN series accessed from Last accessed on July 21, 2014). 39 Looking at the 1969 Census of Agriculture I did not observe a systematic overlapping between grain crop intensity and the presence, or absence, of large oil fields. 54

55 Checks and falsification exercises on the decennial analysis In Table OA1.1 I provide additional analysis on migration rates. Migration rates are constructed with the Census data, by matching information on the CZ of residence five years prior the survey with the CZ of residence at the time of the survey (further details for this match are provided in Online Appendix A2). These results should be interpreted with some caution, as the Census reports migration information with respect to five years prior the Census year only, thus presenting a time-misalignment problem. Moreover, the 1970 Census did not collect information on the place of residence five years before the survey, and therefore I cannot construct out-migration and net migration rates. The estimated coefficients are small in magnitude and imprecise. The results in columns (1), (3) and (4) of Table OA1.1 are estimated only on the first two decades that correspond to the main boom and bust period ( ). In columns (2), (5) and (6) I report only the results for the bust decade (1980s). Results looking at migration rates by skill-level do not indicate differential patterns between the low and the high-skilled population. In Table OA1.2, I make sure that the main results (i.e., tables 2 and 3) are robust to confounding effects and to different specifications of the wage variable. The concern, which I share with other papers in the local labor market literature (e.g., Notowidigdo, 2011), is that the observed migration responses may induce compositional changes that will confound the average effects on wages. The results are robust if instead of the CZ premium I use (i) unadjusted wages; (ii) wages of full-time, full-year workers; (iii) wages of workers who did not migrate in the last five years before the survey. 40 Thus, the incidence analysis of the labor demand shocks is not systematically biased by the endogenously changing characteristics of the local population. In Table OA1.3 I perform further checks described here below. Although large oil fields are present in urban areas (e.g., Los Angeles, Houston, Dallas, Cleveland), they are concentrated more in rural parts of the country. 41 We might expect the economies of metropolitan and rural areas to differ along many dimensions. In panel B, I limit my sample to rural CZs, which I define as those with no MSA as of 1970, to test whether the oil shocks actually affect rural areas only. This leaves me with a sample of 325 CZs. The estimates differ slightly from the baseline estimates in panel A, although the signs and magnitude of the estimated coefficients are largely unchanged. The effects of oil price shocks on nominal wages are larger in magnitude both during the boom (10 percent increase) and the bust (5 percent). Local prices in oil-rich rural areas do not increase as much as in urban areas during the boom (probably reflecting a higher elasticity of housing supply due to larger availability of land), but the drop in prices during the 40 In order to correctly identify in public use Census microdata whether workers migrated from another CZ or just moved within the same CZ I construct a crosswalk based on the migration PUMAs (for 1990 and 2000) and county groups (1980) that maps into CZs. This crosswalk, which is new in the literature, allows also for more detailed analysis of migration rates and it is available upon request. 41 For instance, the metropolitan area of Los Angeles was among the largest world producer of oil in the 1930s from its urban fields, while Dallas and Houston were largely dependent on the oil industry until the early 1990s (??). 55

56 bust is three times larger in magnitude, indicating larger real wage gains during the boom and smaller losses during the bust. These results are consistent with those of Black, McKinnish and Sanders (2005a) who investigate medium/long-run equilibria in the rural region of Appalachia after the coal mining boom of the late 1970s. They find that the responses to local labor demand shocks are larger during the bust than those estimated in this paper, especially in terms of out-migration and wage declines, reflecting even lower economic opportunities in Appalachia outside those provided by the mining sector. They also find smaller spillovers during the boom period with respect to what I find in Table A1. 42 In panel C I report the coefficients for a continuous measure of oil reserves (10 the log of millions of barrel of EUR) interacted with decade dummies. Controlling for a continuous measure of oil reserves aims at estimating the intensity of the treatment as proxied by the amount of oil reserves in large fields in each CZs times the price change in each decade. It is worth noting that the signs are in line with the main specification, and so are the coefficients: the more the oil reserves the larger the impact on CZs outcomes. However, the interpretation of the coefficients is less intuitive and it is not preferred with respect to the baseline specification. In panel D, I exclude all the CZs with any recent discovery of a large oil fields. In seven CZs these discoveries represent the first of a large field in the local area, but they occurred almost exclusively before the major oil price increase in 1973: one occurred in 1970, four in 1971 and two in The results are largely unaffected by excluding these CZs. In panel E, I further test the robustness of the results to the exclusion of the CZs in the major oil producer state, Texas. The results are robust to this specification indicating that oil booms and busts have similar effects across regions, and are not peculiar to one state. In panel F, I exclude the top four CZs in terms of oil reserves to make sure that the results are not driven by outliers, and it does not appear to be the case. Finally, in panel G, I allow for less conservative standard errors clustered at the CZ-level, and as expected the standard errors generally decrease indicating correlated errors at the state level. 42 Similarly, Marchand (2012) analyzes the direct employment effects of oil booms and busts on the oil extraction industry in Alberta, Canada, and the multiplier effects into other local sectors. He finds small spillovers and he does not investigate the effects of the bust on the local economy. 56

57 Online Appendix OA1: Figures Figure OA1.1. CRs to extreme positive, extreme negative and normal oil price shocks Panel A. Population (log) Figure OA1.2. CRs of local labor markets outcomes to grain shocks Panel A. Population (log) Panel B. Total employment (log) Panel B. Total employment (log) Panel C. Annual earnings per worker (log) Panel C. Annual earnings per worker (log) timated by local projections (see equation (2)). The dark line Note: The figure plots separately for each outcome the Cumulated Response functions (CRs) over a 5 years horizon esti- plots the CR to a positive grain price shock, while they grey line plots the CR to a negative grain price shock as described mated by local projections. Differently from equation (2) and in this Appendix. The shaded areas represent 95 percent pointwise confidence intervals robust to state clustering. Figure 5, I now estimate the relative oil-rich versus non-oil CZs CRs to Extreme Positive (blue), Extreme Negative (red), 57 and Normal (grey) oil shocks defined as the top, bottom and middle tercile of the log oil price distribution. The shaded areas represent 95 percent pointwise confidence intervals robust to state clustering. Note: The figure plots separately for each outcome the Cumulated Response functions (CRs) over a 5 years horizon es-