Energy Economics 59 (2016) 81–95
Contents lists available at ScienceDirect
Energy Economics
journal homepage: www.elsevier.com/locate/eneeco
Economics of modern energy boomtowns: Do oil and gas shocks differ
from shocks in the rest of the economy?☆
Alexandra Tsvetkova ⁎, Mark D. Partridge
Ohio State University, Columbus, OH, USA
a r t i c l e
i n f o
Article history:
Received 8 August 2015
Received in revised form 23 June 2016
Accepted 16 July 2016
Available online 09 August 2016
JEL classification:
O13
Q33
R11
Keywords:
Employment growth
Job multipliers
Energy boom effects
Instrumental variable approach
a b s t r a c t
The US shale boom has intensified interest in how the expanding oil and gas sector affects local economic performance. Research has produced mixed results and has not compared how energy shocks differ from equal-sized
shocks elsewhere in the economy. What emerges is that the estimated impacts of energy development vary by
region, empirical methodology, as well as the time horizon that is considered. This paper captures these dimensions to present a more complete picture of energy boomtowns. Utilizing US county data, we estimate the effects
of changes in oil and gas extraction employment on total employment growth as well as growth by sector. We
compare this to the effects of equal-sized shocks in the rest of the economy to assess whether energy booms
are inherently different. The analysis is performed separately for nonmetropolitan and metropolitan counties
using instrumental variables. We difference over 1-, 3-, 6-, and 10-year time periods to account for countyfixed effects and to assess responses across different time horizons. The results show that in nonmetro counties,
energy sector multiplier effects on total county employment first increase up to 6-year horizons and then decline
for 10-year horizons. We also observe positive spillovers to the non-traded goods sector, while spillovers are
small or negative for traded goods. In metro counties, there are no significant effects on total employment, although positive spillovers are present in some sectors. Yet, equal-sized shocks in the rest of the economy produce
more jobs on average than oil and gas shocks, suggesting that policymakers should seek more diversified
development.
© 2016 Elsevier B.V. All rights reserved.
1. Introduction
The recent shale boom in the US has intensified scholarly interest in
how expansion of the resource extraction sector affects local economic
performance of impacted areas. Although international research mostly
documents the so-called resource curse (an inverse relationship
between resource sector dependence and long-run economic performance), less has been done at the subnational level in developed countries, though there is increasing interest (Freeman, 2009; Gylfason et al.,
1999; James and Aadland, 2011; Papyrakis and Gerlagh, 2007; Sachs
and Warner, 1999, 2001). For example, recent research in the US context often looks at how a growing energy sector affects short-term
local economic performance during the previous boom decade
(Brown, 2014; Weber, 2012, 2014; Weinstein, 2014). It remains to be
seen if localities that benefit from resource extraction now will be able
☆ The authors acknowledge and appreciate the support of the USDA AFRI grant
#11400612 titled “Maximizing the Gains of Old and New Energy Development for
America's Rural Communities.”
⁎ Corresponding author at: Agricultural Administration Bldg., 2120 Fyffe Rd., Columbus,
OH 43210, USA.
E-mail addresses: [email protected] (A. Tsvetkova), [email protected]
(M.D. Partridge).
http://dx.doi.org/10.1016/j.eneco.2016.07.015
0140-9883/© 2016 Elsevier B.V. All rights reserved.
to retain some of the gains in a prolonged bust period. What is clear is
that the conclusions of past research largely depend on the specific
settings of the study, such as time frame, sample, etc. (Munasib and
Rickman, 2015).1
Most subnational research within the US has focused on selected
regions (Haggerty et al., 2014; Michaels, 2011; Paredes et al., 2015) or
only nonmetropolitan areas (Brown, 2014; Munasib and Rickman,
2015; Weber, 2014). Detailed studies of the whole country are rare
(one exception is Weinstein (2014)). Academic investigations of the
energy sector’s employment effects usually concentrate on total employment or just a few sectors (Brown, 2014; Weinstein, 2014). Besides,
the existing studies generally fail to explicitly account for the nonuniform temporal association between energy sector and changes in
employment growth.
This paper contributes to the literature by presenting what we believe is the most detailed account of the effects of energy sector on US
local employment. We separately consider metropolitan and nonmetropolitan areas by examining total employment as well as employment in
1
The academic studies helped discredit many so-called “economic impact” studies that
find very large spillovers from energy development, most of which were funded by the energy industry (Kinnaman, 2011).
82
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
14 sectors to assess how energy booms differentially affect regional
economies. We also examine the degree to which our analysis is affected by regional heterogeneities. Our time frame is 1993 to 2013, although most of the analysis takes place after 2000. Tracing the effects
through 2013 is crucial for capturing the dynamics of the last decade,
which is characterized by rapid expansion in extracting unconventional
oil and gas sources. Our empirical strategy also examines shocks of various durations to get a clearer picture of how the effects of energy
booms evolve over time as supply chains and displacement effects
take various amount of time to work their way through regional economies. Finally, we are the first to compare oil and gas shocks to equalsized shocks to assess whether and how energy booms differ from
growth in the rest of the economy.
The estimation is based on an instrumental variable (IV) technique
and first-differencing over 1, 3, 6, and 10 years. We employ two combinations of instruments that are based on geological measures of oil and
gas resources as well as historical measures of drilling activity. The
differencing approach removes location-fixed effects. The IV accounts
for possible endogeneity that might be associated with the location of
the development that relate to factors such as attitudes of the voters,
the regulatory framework, or long-run expectations of economic conditions (e.g., desperately poor economies looking for jobs or pro-business
environments). There are also time-varying factors that might affect
when such development takes place (e.g., technological improvements,
changes in world energy prices, or business cycle effects might change
willingness of a locality to accept development). The IV approach also
mitigates against potential data measurement issues.
The results suggest that energy sector growth measured by oil and
gas employment has positive net spillovers to a number of other sectors
in nonmetropolitan areas. In traded goods sectors, however, there is
some evidence of crowding out effects. In metropolitan counties we observe no statistical link for total employment, although the results at the
sectoral level show some positive spillovers. Yet, energy shocks tend to
have smaller total employment effects than equal-sized standard
shocks across the rest of the economy, suggesting that energy has
fewer complementarities with other sectors than average. Likewise,
population responses to both energy shocks and representative shocks
elsewhere in the economy are both small, suggesting limitations in
the role of agglomeration economies supporting long-term growth in
modern booms. Overall, given that the job effects of energy booms are
relatively modest in magnitude compared to the scale of the rest of
the economy, local economies would be better off if they were to experience broad-based growth rather than energy booms both in terms of
multiplier effects and enhanced economic diversity (with less risk for
a possible natural resource curse).
The rest of the paper is organized as follows. The next section reviews literature with specific focus on the possible mechanisms of the
resource curse and on the subnational studies within the US. Section 3
presents empirical methodology and the data sources followed by the
discussion of results in Section 4. Section 5 briefly covers various sensitivity checks, whereas Section 6 contains concluding remarks.
2. Literature review
The literature that investigates the relationship between resource
endowment and various measures of socioeconomic performance and
well-being at the national level has a long tradition. It often documents
a negative association between resource dependence and long-run
growth (Gylfason et al., 1999; Sachs and Warner, 1999, 2001;
Sala-I-Martin et al., 2004), which is commonly referred to as the resource curse. Although some authors find stimulating effect of resource
abundance on growth (Alexeev and Conrad, 2009; Cavalcanti et al.,
2011), others argue that sluggish economic performance may stem
from market imperfections that are unrelated to resources (Gylfason
and Zoega, 2014; Manzano and Rigobon, 2001) or that the findings in
the resource curse tradition are driven by the way the main variables
are measured (Brunnschweiler and Bulte, 2008), the evidence of the
negative impacts of natural endowments on economic performance is
somewhat more plentiful.
Existing research offers about a dozen explanations at both macroeconomic and microeconomic levels to the negative relationship within
the resource curse perspective. Frankel (2010), in his detailed nationallevel survey, lists six potential mechanisms of the resource curse. Perhaps the most general ones are the tendency of resource-rich countries
to be war-ridden and the possibility of a universal decline in the world
markets for resources. Both these factors can plausibly dampen growth
rates in the resource-dependent economies and make them more vulnerable overall. An alternative (or additional) explanation is a volatility
of resource prices in the world markets, which may lead to macroeconomic instability and excessive governmental spending with a potential
to harm growth. Over time, however, the negative effects of such a decline and/or volatility should erode if an economy effectively adjusts
to the changes in the global markets.
It also seems that resource-rich nations tend, on average, to lack the
ability to modernize their economies and institutions (Mehlum et al.,
2006; Ross, 1999). This may be intrinsically related to other possible
mechanisms of the resource curse, namely corruption, elitism and
low-quality institutions (Bhattacharyya and Hodler, 2010; Bjorvatn
et al., 2012; Sala-i-Martin and Subramanian, 2013), which are often
present if a government or hereditary elites possess the physical command of the resources. Another widely studied mechanism is the
“Dutch Disease,” which basically means crowding out manufacturing
(or other tradable sectors) due to increasing rents and wages in the extraction sector (Sala-i-Martin and Subramanian, 2013). Other explanations (in the context of both nations and subnational regions) revolve
around resource-abundant regions’ failure to diversify their economies
and export structures (Murshed and Serino, 2011); lower educational
attainment and human capital accumulation (Black et al., 2005; Blanco
and Grier, 2012; Gylfason, 2001), and crowding out entrepreneurship
and innovation (Betz et al., 2015; Gylfason, 2000; Sachs and Warner,
1999). See Van der Ploeg (2011) for a more comprehensive review.
Recent research on the effects of energy sector has increasingly used
subnational data, which reduces the influence of unobserveables such
as institutional quality, though there is still heterogeneity across different subnational regions (Fleming and Measham, 2014; Munasib and
Rickman, 2015). The expected overall positive link between endowment with unconventional energy resources and employment and income is not straightforward if one takes into account extensive use of
long-distance commuters (Kinnaman, 2011; Munasib and Rickman,
2015), and potential adverse health and environmental effects
(Sovacool, 2014) that possibly dampen property values in extraction
areas (Boxall et al., 2005). Indeed, the empirical evidence on the impact
of resource abundance on regional US socioeconomic outcomes is
mixed and may depend on the level of aggregation.
At a state level, the evidence tends to point to the negative association much in line with what has been observed in the international settings, although there are exceptions. Based on cross-sectional analysis,
Papyrakis and Gerlagh (2007) conclude that resource dependence measured by the share of primary-sector gross state product (GSP) output
does not directly slow growth. Instead, large primary sector is related
to lower investments, schooling, openness, research and development
(R&D), and higher corruption, which, in turn, translate into sluggish
growth. Using panel data, Boyce and Emery (2011) find a negative association between resource abundance and growth rates but a positive effect on income levels. Freeman (2009) finds that intensity of
employment in the primary sector (mining and agriculture) is negatively related to US GSP per capita growth between 1977 and 2002. James
and James (2011) explain the observed negative relationship between
resources and economic growth by slower growth in mining compared
to other sectors. When this factor is accounted for in their models, the
results suggest that more resource-dependent states enjoyed greater
per capita personal income growth between 1980 and 2000.
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
At a county level, in contrast, the empirical analyses of the recent energy boom usually reveal a stimulating effect of energy endowments.
Paredes et al. (2015) analyse counties in the states of New York and
Pennsylvania using a ban on shale gas fracking imposed in New York
state as a natural experiment. Their results suggest that unconventional
gas wells have a positive contemporaneous effect on income and employment, although the coefficients in the former case are rather
small. Weber (2012) estimates the impact of gas production value on
employment, income and poverty for counties in the states of Colorado,
Texas, and Wyoming using difference-in-difference and instrumental
variable approaches. He concludes that shale gas extraction modestly
contributed to employment and income but did not affect poverty
rates. Weber (2014) focuses on nonmetropolitan counties in the states
of Texas, Louisiana, Arkansas, and Oklahoma during the 2000s. According to this analysis, each gas-related mining job was associated with additional 1.4 non-mining jobs. In a comparable analysis that covers a
nine-state region of the US Great Plains and Oil Patch (the states
of Arkansas, Colorado, Kansas, Louisiana, Nebraska, New Mexico,
Oklahoma, Texas, and Wyoming), Brown (2014) finds a smaller increase in employment per each additional gas-related mining job (0.7
vs. 1.4 in the previously cited paper). Counties that increased gas production saw small employment gains in construction, transportation,
and services, while retail trade and manufacturing were not affected.
Based on the analysis of all counties in the lower 48 states over the
years 2001–2011, Weinstein (2014) reports an employment multiplier
from shale boom-related labour market restructuring of approximately
1.3. All studies considered used different time periods as well as samples
and measurement of the crucial variables; some used production to
proxy employment and then tried to extrapolate employment even
though production and employment are not necessarily highly linked.2
There is some evidence, though, that the link between resources and
socioeconomic outcomes is not uniform across the country. Munasib
and Rickman (2015) find that unconventional oil and gas production
in the state of North Dakota has large and statistically significant positive effect on total employment and employment in retail, accommodation, and construction. For Arkansas, positive effects are observable only
in the most extraction-intensive counties, while they find no impact for
the state of Pennsylvania.
County-level investigations that examine previous energy boom–
bust cycles present somewhat ambiguous results. On the one hand,
James and Aadland (2011) show that personal income growth between
1980 and 1995 was slower in counties with a larger primary sector, thus
lending support to the natural resource curse hypothesis. On the other
hand, Michaels’ (2011) investigation of resource abundance effects on
economic performance of counties in the US South during the 1940–
1990 period indicate that areas situated over oilfields enjoyed higher
employment, income, and population density, although the effects lessened after about 1960. In the county-level analysis of more recent years,
Peach and Starbuck (2011) find small but positive effect of oil and gas
production value on employment and income in the state of New
Mexico in 1960, 1970, 1980, 1990, and 2000. In a study dating to
1960, Allcott and Keniston (2014) find that oil- and gas-endowed (in
terms of reserves) counties experience employment growth during energy booms, although these effects quickly reverse in busts. Analysis of
2
We downloaded the 2000–2011 oil and gas production data from the USDA Economic
Research Service, which has a little different coverage in years than our EMSI employment
data (http://www.ers.usda.gov/data-products/county-level-oil-and-gas-production-inthe-us.aspx). For the years that overlap, we regress the percent change in EMSI oil and
gas production employment on the percent change in oil production and the percent
change in natural gas production (this is the closest we can make our data correspond
to the production data for a comparison). The R2 statistics for the 1-, 3-, and 6-year percent
changes are .0001, .0005, and .0006, respectively, for the nonmetro sample. The corresponding metro sample R2 statistics are .0000, .1043, and .4044. Thus, the correlations
are very low, especially for the nonmetro sample, suggesting that one should be very cautious when interpreting employment multipliers that are indirectly computed with production data.
83
income and employment from 1969 to 1998 by Jacobsen and Parker
(2014) reveals that oil and gas extracting counties enjoyed job growth,
particularly in mining and non-tradable sectors, during the boom but
suffered larger income declines and greater unemployment in the
bust. Somewhat different findings come from Haggerty et al. (2014),
who fail to find significant relationship between county specialization
in oil and gas, as well as duration of such specialization, on change in
employment between 1980 and 2011 in the US West.
3. Estimation approach, variables, and data sources
The impacts of changes in resource industry are likely to depend on
the specific setting and time horizon. In this paper, we are interested in
two important policy indicators, employment and population. We do
not consider income because of theoretical concerns regarding interpretation in a spatial equilibrium framework and most directly, in an employment growth model, as income/wages are jointly determined
with employment.3 In contrast to recent research that focuses on nonmetro areas (Brown, 2014; Munasib and Rickman, 2015; Weber,
2014), we separately study metropolitan and nonmetropolitan areas.4
While extraction activity predominantly takes place in rural areas, the
use of long-distance workers and tendency of energy companies to locate their headquarters in metropolitan areas suggest the need for
such a distinction. Existing research usually concentrates on total employment (Paredes et al., 2015; Weber, 2012, 2014) or considers just a
few sectors (Brown, 2014; Weinstein, 2014). We take a step further
and trace the effects at a more disaggregated level across 14 sectors.5
Various parts of the regional economy are likely to respond differently
to energy sector expansion. Whereas positive effects can be expected
in accommodation or construction, traded goods may be crowded out
as suggested by the Dutch Disease literature.
The time path of an energy boom is likely to vary over time as new
supply chains are established and complementary sectors are built
out, whereas other sectors and activities are crowded out. To capture
the dynamic time path of energy boom, we separately employ differencing of the dependent and main explanatory variables over 1-, 3-, 6-, and
10-year periods. This allows us to assess the dynamics of a boom in
which the multiplier effects and offsetting displacement effects work
at a different pace through the supply chains across different industries
and county types. The time span of the study is from 2001 to 2013 for
the former three differences and from 1993 to 2013 for the 10-year difference. Extending the analysis to 2013 is important for capturing recent
trends in unconventional oil and gas extraction and makes this examination one of the most current in the literature. A potential weakness
of this study (and other recent studies) is our inability to comment on
longer-run employment effects of energy sector expansion because
there was no sustained bust period over the 2000–2013 period. In
terms of total job growth, all we observe is the (net) positive labour
3
The standard model used to assess US regional trends is the spatial equilibrium model
that assumes highly mobile factors of production and profit and utility maximization
(Glaeser and Gottlieb, 2008), which result in utility and profits being equalized across
space. In this case, people trade-off income with site-specific amenities (weather, landscape, etc.) so that income by itself is not necessarily a measure of well-being. In mining
boomtowns with socioeconomic, congestion, remoteness, and environmental concerns,
high income is to a large extent a compensating differential, while during the bust, some
of the income decline reflects less of a need for such compensating differentials. In a
cross-country international comparison of (say) natural resource curse, income differences are more meaningful.
4
We use the December 2003 Office of Management and Budget (OMB) metropolitan
area definitions based on the 2000 Census.
5
The specific sectors are: agriculture; manufacturing; construction; transportation and
warehousing; retail trade; wholesale trade; accommodation and food services; real estate,
finance and insurance; information services; professional, scientific and technical services;
tradable industries; non-tradable industries; and industries that are “upstream” to oil and
gas extraction. The definition of all sectors, except for tradable, non-tradable and “upstream” industries, follows standard NAICS2012 industry classification. The description
of tradable, non-tradable and “upstream” industries is in Appendix A.
84
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
demand shock and we cannot average over both the boom and inevitable bust.
Our empirical approach has several advantages. First, differencing
over varying time periods removes the unobservable county-fixed effects. Using differencing also offers a nice advantage over panel or within models that use fixed effects because we can control for persistent
disequilibrium level effects that may affect growth even after differencing out fixed effects (described further below).6 In particular, we can explicitly control for the share of historical mining employment at the
time of the previous boom–bust cycle to directly assess whether historical factors related to infrastructure development affect contemporaneous oil and gas development spillovers, which we will see plays a role
in the results. The standard errors are clustered within US Bureau
of Economic Analysis (BEA) economic areas to account for spatial
correlation of the residuals within these particular economic regions.
In sensitivity analysis, we also clustered the errors by county to account
for serial correlation of the residuals, but this did not tangibly affect the
results. Given the more prominent concern with possible spatial
dependence in county-level data, we cluster by BEA region, though
the other results are available on request.
The second advantage is instrumenting for relative energy employment growth with variables that reflect (1) historic energy production
(because shale is often located in counties with a conventional oil and
gas production history and these places may have more supporting infrastructure) and (2) geological conditions that drive the resource’s
availability and allowing the instruments to vary with time. Such an estimation approach follows from a possibility that the relative profitability of energy development may be location-specific and time-varying.
For example, struggling places or places with pro-business political
leadership may welcome energy development, which can change over
time with local cyclical conditions. More details about the instruments
are provided below.
The third advantage of our approach is that we use actual energy
employment using a proprietary data set described below. In our setting, it is energy employment that directly affects the labour market
outcomes of interest. Most of the previous literature used some measures of reserves (which is an instrument in our case) or production
data, mostly due to limited data availability. Yet, the first-order effects
of the energy resource/reserves on a local economy occur indirectly
through their impact on local energy employment. If there is little or
no related employment, reserves have a much smaller impact on
outcomes—e.g., the shale resource has long been under ground, but
did not have any tangible effects until development. Oil and gas production data may be a less suitable measure as it fails to capture employment changes prior to production or as production matures. As Kelsey
et al. (2016) note, well drilling and the building of associated infrastructure such as pipelines accounts for a disproportionately large share of
new jobs created within the expanding energy sector. Many jobs,
however, are created before production starts. After production fully
matures, energy sector employment needs are much smaller.
Eq. (1) presents the empirical specification
ΔY ic ¼ β0 þ β1 ΔEnVar c þ β2 ΔDShockc þ Xβ þ θt þ εc
ð1Þ
6
See Wooldridge (2010, Chapters 10 and 11) for more details of the relative tradeoffs in
using fixed effects versus first differencing in meeting the statistical assumptions. Both approaches have their relative merits. In our case, aside from being forced to omit the time
invariant variables we use in the differenced models, the 6- and 10-fixed effect models
would produce exactly the same results as the differenced models because there are only
two periods being considered (Wooldridge, 2010). Nonetheless, we did experiments
using fixed effects models in the 1- and 3-year models, which forced us to drop
the time-invariant variables. Yet, the instruments were very weak with first-stage
F-statistics always being below 3 and we did not consider these models further. Angrist
and Pischke (2008, 2014) and Wooldridge (2013, Appendix 13A and 14A) note that the
main empirical concern with first differencing is that there is typically a need to cluster
by location (in spatial data) to account for serial correlation, which would also apply for
fixed effects models if there is serial correlation in those specifications.
where ΔYic =Yict −Yict−n ; ΔEnVarc =EnVarct −EnVarct−n and ΔDShockc =
DShockct − DShockct − n; subscript i denotes a sector or total employment, c refers to a county, t is a time period with n=1, 3, 6, 10, which
represents the differencing used for each set of models.
The first component of the dependent variable ΔYic (Yict) is the
change in total employment (or employment in a sector) between
two periods (separated by 1, 3, 6, or 10 years) divided by the total
county employment in the base period. The second element of the
dependent variable ΔYic (Yict−n) is calculated in the same way for a preceding period7. The difference between the two values is the dependent
variable. In the results tables below, we refer to the dependent variable
as “the change in employment growth rates between periods” for brevity. In calculating the dependent variable for sectors, for each element
Yic, we divide the change in employment between periods in a given
sector by total employment in a county in the base period in order to
scale the sector’s growth to the total size of the economy. This allows
us to directly calculate multipliers because in this case they are scaled
to total employment and they are also comparable across models.
We define oil and gas employment (energy employment) as the
sum of employment in two industries, NAICS2111 (oil and gas extraction) and NAICS2131 (support activities for mining). The change in energy employment growth, ΔEnVarc (or DiffEnVar in tables) over the four
different time spans is the main explanatory variable. We first calculate
EnVarct as the difference in total oil and gas employment in a county between the end and beginning periods divided by the initial total county
employment, producing the contribution of energy to total employment
growth. Differencing EnVar over 1-, 3-, 6-, and 10-year periods produces
DiffEnVar. Because both the energy explanatory variable and dependent
variables are divided by total employment, the value of β1 coefficient is
interpreted as a multiplier. For example, when modelling total employment growth as the dependent variable, a β1 that equals 1 means that
total employment changes by exactly the same number of jobs (or
percentage) as energy employment, or there are no net spillovers of
energy employment on other sectors on balance. If β1 is greater (less)
than one, there are net positive (negative) spillovers on the other
sectors.8
One key control variable is the industry mix measure DShockct that
reflects demand shocks in all other industries in a county.9 We calculate
DShock as follows.
DShock ct ¼
X
Scit−n NGit ;
ð2Þ
i
where Scit − n stands for the employment share of industry i
(i ≠ NAICS2111 or NAICS2131) in county c in the beginning period and
NGit refers to the national employment growth rate in industry i
between years t − n and t (n = 1, 3, 6, 10). The industry mix term represents the predicted growth rate of the county’s employment if all of its
7
In the 10-year differenced models for total employment, the dependent variable is calculated as employment growth rate between 2003 and 2013 minus employment growth
rate between 1993 and 2003;in the 6-year differenced analysis for total employment, the
dependent variable is calculated as employment growth rate between 2007 and 2013 minus employment growth rate between 2001 and 2007; in the 3-year differenced analysis
for total employment, the dependent variable is calculated as employment growth rate
between 2010 and 2013 minus employment growth rate between 2007 and 2010; employment growth rate between 2007 and 2010 minus employment growth rate between
2004 and 2007; employment growth rate between 2004 and 2007 minus employment
growth rate between 2001 and 2004; in the 1-year differenced models for total employment, the dependent variable is calculated in a similar fashion: employment growth rate
between 2012 and 2013 minus employment growth rate between 2011 and 2012, etc.
In the analysis of sectors, the dependent variables are calculated in the same way with
the difference between two periods divided by total county employment in the base period so that the corresponding regression coefficients are consistently reported as
multipliers—i.e., they are the growth rate in total county employment due to that sector.
8
See, for example, Marchand (2012) for a comparable approach. Our coefficients
are more direct and are easier to interpret because we do not use a logarithmic
transformation.
9
This measure is sometimes called the Bartik (1991) instrument because it is assumed
to be exogenous.
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
industries (minus oil and gas) are growing at the national growth rate.
This variable is typically presumed exogenous to the modelled relationships, as it utilizes initial industry composition of a county and projects
county growth based on the national growth rates that are not influenced by growth dynamics of a single county. Controlling for aggregate
shocks also accounts for factors that may be correlated with new energy
development and may bias the coefficients if omitted. The oil and gas
sector is omitted from the industry mix calculation so that we can compare the size of the oil and gas variable coefficient to the industry mix
coefficient to ascertain whether an energy shock has a different effect
compared to an equally sized typical shock outside of the oil and gas
sector. In Eq. (1), ΔDShockc (DiffDShock) is measured at 1-year, 3-year,
6-year, and 10-year differences depending on the model.
Our models also include controls for initial-period and deep-lags of
socioeconomic conditions X that are widely used in the literature, in
which the lags mitigate any endogeneity that might be present. First,
to account for a historical agglomeration effect, the natural log of county
population in 1980 is included. We choose 1980 as a year during the
peak of the previous US energy boom and thus we control for agglomeration features that may be correlated with energy development (including possible energy infrastructure) and future population growth. We
also include lagged educational attainment to account for persistent
human capital effects (share of adult population with only high school
diploma and the share of adult population with at least four years of college in 1990).
To account for industry composition effects that may affect growth,
we include 1990 industry employment shares in manufacturing and agriculture; lagging these variables to 1990 helps avoid any endogeneity
due to sorting in anticipation of the shale boom. We also control for
the 1985 mining employment share because accounting for the historical size of mining as the 1970s–1980s energy boom came to an end captures any historical energy sector agglomerations that may in turn
influence the contemporaneous size of the industry. These include mining and energy legacy effects that are related to the presence of energy
infrastructure such as pipelines and service companies or a tradition of a
more favorable climate for energy development. Finally, we include
time period dummies, θt.10
As noted earlier, local energy employment may be related not only
to energy endowments but also to other factors that possibly influence
the dependent variables—e.g., environmentally sensitive residents may
object to resource extraction due to its negative quality-of-life effects.
To avoid potential endogeneity, we follow other studies (Brown,
2014; Weber, 2012, 2014; Weinstein, 2014) and employ an instrumental variable (IV) approach based on geological measures to complement
our first difference approach. Unlike previous studies that use only one
instrument and estimate exactly identified models, we try multiple instrument combinations to assess their relative merits. A key advantage
of using several instruments is our ability to test the validity of the
over-identification assumptions rather than simply assuming they
hold.11 Using IV also alleviates potential measurement error problems
in the oil and gas employment data and accounts for omitted variable
bias including the omission of spatial lags of the dependent or explanatory variables if there is spatial dependence.
We use higher-quality employment data compared to what have
been used in the past. The past work often uses mining employment
(not oil and gas) and typically extrapolates due to numerous observations at a county level suppressed for confidentiality.12 Yet, the data
we use may still have measurement issues. For one, firms and their
10
We also tried the 1990 share in oil and gas employment to capture long-term legacy
effects of oil and gas but the results are very similar to the 1985 mining share.
11
A popular instrument has been whether the county sits on shale deposits (Brown,
2014; Weber, 2012), which would mostly be useful for predicting unconventional oil
and gas drilling, whereas our multiple measures also help us identify conventional drilling
as well.
12
Alternative data used in previous work are place-of-residence Census data, which
have their own limitations when used to study employment change at a county level.
85
workers commuting for work to a nearby county could be counted as
working in their place-of-residence county if the firm is based there.
Also, a surge in new jobs that lasts a short period of time or happens
at the end of a calendar year might affect the validity of an annual employment average, though this would affect all studies using annual
data.
We develop five instruments related to tight oil and shale gas geological abundance of the resource as well as the location’s historical legacy of drilling that might affect the available infrastructure such as
pipelines. They are: (1) shale oil reserves of the county at the beginning
of the decade; (2) beginning of the decade shale gas reserves in the
county; (3) the county’s number of square miles with oil and gas
wells in the 1980s; (4) following Weber (2012), the percentage of the
county’s territory over shale plays; and (5) average thickness of the
shale in each “shale play” if the county sits above a play, as thickness
is a proxy for resource abundance.13 The instruments are further described in Appendix B. The geological instruments reflect the exogenous
geology or predetermined availability of reserves while drilling intensity in the 1980s reflects whether there is a history of production in the
county (after conditioning on mining employment after the early
1980s bust). That could also suggest conventional reserves that could
be coaxed out with modern technology or suggest better shale reserves
given the association of shale and conventional reserves.
Following the econometric approach employed by Duranton and
Turner (2011), we assume that after conditioning on the size of the
1985 mining sector and other lagged variables described below, the
geological variables only indirectly influence contemporaneous energy
employment growth by affecting the size of the energy sector employment share, which is the definition of a good instrument. Similarly, conditioning on 1980s population would further account for long-term
agglomeration effects that might be correlated with current energy development and contemporaneous job growth.
A practical concern stemming from using multiple observations per
county is that the geological/historical scale instruments are fixed over
time and might appear to provide no more information than simply
considering fixed effects or first differences. Yet, that is not true because
shale energy technology evolves over time, suggesting that geological
reserves are more important in later time periods with improved technology. Likewise, commodity prices are set on national and world markets and are not affected by one US county’s production, so year-to-year
variation in the price of oil and natural gas also change the value of the
resource and whether extraction is profitable. We include a series of instruments with time interactions to account for the time-varying
endogeneity.
Given the large number of instrument combinations, after a series of
diagnostic tests, two combinations of instruments (and their time interactions) were selected based on their strength in the first stage and
whether they pass over-identification tests for the total and sectoral
employment models. Instrument I (IV I) includes two measures that approximate thickness of oil and gas shale deposits and the oil and gas
drilling intensity in the 1980s at the county level. Then for further
check of robustness, instrument II (IV II) supplements IV I with
measures of recoverable shale oil and recoverable shale gas reserves.
Appendix B describes how instrument components are determined.
We estimate Eq. (1) using 2SLS. Thus, the identification strategy relies
on first-differences to eliminate county-level fixed effects of various duration and IV using exogenous (or at least deep-lagged predetermined)
instruments to account for any time-varying endogeneity, while
accounting for industrial shocks and historical measures to capture
long-term agglomeration effects that might impact contemporary
drilling.
13
Thickness is associated with how much recoverable resource is in the shale—e.g., see
EIA’s reports on South Africa and Spain at https://www.eia.gov/analysis/studies/
worldshalegas/pdf/South_Africa_2013.pdf and https://www.eia.gov/analysis/studies/
worldshalegas/pdf/Spain_2013.pdf.
86
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
The primary data source is a proprietary dataset acquired from
Economic Modelling Specialists Intl. (EMSI)14 with information on
county employment disaggregated at 4-digit NAICS level. EMSI data
have been successfully used in recent studies.15 The dataset allows us
to measure oil- and gas-extraction employment, as well as other sectors
of interest, more precisely (especially in calculating our energy and
industry mix terms). By contrast, detailed employment data from secondary government sources are incomplete because of suppression
for confidentiality.16 For example, for the two key energy industries,
NAICS 2111 and 2131, the Quarterly Census of Employment and
Wages (QCEW) only reports employment for about 20 percent of all
counties, though these counties account for about 87 percent of national
oil and gas employment. The correlations between unsuppressed values
of NAICS2111 and NAICS2131 employment reported by the QCEW for
year 2013 and the corresponding values from EMSI data are, respectively, 1.0 and .99999. This suggests that EMSI data are quite accurate and, in
fact, nearly perfect for the vast majority of oil and gas employment in
the country.17 Even if EMSI data were less accurate for the remaining
counties, the counties for which the QCEW suppresses values are typically small in terms of oil and gas employment and should not make
much difference given their low “leverage” in the regression (especially
in difference in growth rate form that we employ for the dependent and
the main explanatory variables). To the extent that there is a measurement error, our IV approach should adjust for this issue. For other variables, we supplement the EMSI data with information from the US
Census Bureau, the US Energy Information Administration and other
publicly available sources described below.
The dependent variables, energy measures, industry mix, and employment shares in manufacturing and agriculture are calculated from
the EMSI data. All shale-related indicators are from the geospatial files
provided by the US Energy Information Administration (EIA).18 The
data on the 1980s drilling intensity is aggregated from the US Geological
Survey geospatial files.19 The US Census Bureau is the source for population and educational attainment. The 1985 mining share variable is
derived from the US Bureau of Economic Analysis data using the
methods described in Partridge and Rickman (2006).
4. Results and discussion
We first separately describe the total employment results in nonmetropolitan and metropolitan counties (Table 1) followed by discussion of population response and selected sectoral results. All tables
include the four different lengths for the time differencing and both instrument sets, whereas Appendix C has descriptive statistics. All models
include controls for education, initial economic conditions and time
periods in addition to the displayed variables unless indicated otherwise. Since our estimation relies on an IV approach and uses multiple
variables and their time interactions as instruments, this may increase
collinearity and reduce the first-stage strength of the instruments. This
14
www.economicmodeling.com.
Examples include Betz et al. (2015), Dorfman et al. (2011), Fallah et al. (2011), Fallah
et al. (2014), Nolan et al. (2011).
16
EMSI uses several government sources such as the Bureau of Economic Analysis REIS
data, County Business Patterns, and Quarterly Census of Employment and Wages (QCEW).
They also use extrapolation programs to ensure that for every county, summing all the individual industry employments equals the county total employment and that summing an
industry’s employment across all counties equals the state total industry employment.
Dorfman et al. (2011) provide more details of EMSI’s process.
17
The correlation should not necessarily be perfect because the QCEW does not include
self-employed while EMSI includes self-employed using data from the American Community Survey (ACS). The ACS reports people who are self-employed if their primary job is
self-employment in contrast to the BEA self-employment estimates that include all forms
of self-employment (partial or full). This naturally translates into much larger share of selfemployed in a local economy according to the BEA data.
18
http://www.eia.gov/pub/oil_gas/natural_gas/analysis_publications/maps/maps.
htm#geodata.
19
http://certmapper.cr.usgs.gov/geoportal/catalog/search/resource/details.page?
uuid=%7B985D2AB7-B159-46C2-BD58-99EC3E366C4F%7D.
15
does not appear to be the case as evidenced by the F-statistic on the excluded instruments in the first stage (a measure of instrument strength)
well above 1020 in most cases.
4.1. Results for total employment
Before turning to the main results in Table 1, the Durbin test results
at the bottom of the table suggest that the null hypothesis that the energy variable is exogenous can be rejected in four of the eight nonmetro
models and it is never rejected in the metro models. Given that energy is
such a small share of metropolitan economies, the latter result is not
surprising. Though these results suggest we could use OLS, we report
the 2SLS results here given the dominance of this approach in the literature. We summarize the OLS results for total employment in the Sensitivity Analysis section and OLS results for the tradable and non-tradable
sectors in Appendix D. As we discuss, the OLS and IV results are quite
consistent, so the conclusions are economically equivalent regardless
of one’s preference for OLS or IV.
Using the rule of thumb that the first-stage F-test should be greater
than 10, the instruments are quite strong with the exception of the
two cases in metro models and two cases in the nonmetropolitan
ones. The Sargan over-identification test p-values (Sargan, 1958)
imply that some of the equations may not be well identified, especially
the metro models.21 This calls for caution in interpreting the results, although the geological nature of the instruments and the unanticipated
nature of the shale revolution (and global energy prices), as well as
differencing out fixed effects, suggest our empirical approach is sound.22
We are primarily interested in the results for the nonmetro counties
because the shale revolution would have the largest relative impacts in
less-populated counties. The change in the energy employment variable
(DiffEnVar) is the main explanatory variable. The nonmetropolitan results in columns [1] and [2] of Table 1 suggest that expanding energy
sector employment is associated with net positive spillovers to other
sectors. For example, the 1.261 coefficient in column [1] suggests that
over the course of one year, for every 100 direct jobs created in the oil
and gas sector, another 26.1 jobs are indirectly expected elsewhere in
the local economy. In some sense, these are consistent with Brown
(2014), who hypothesized that thin labour markets in rural areas may
have less slack for expanded employment, which would seem to especially apply in the short-run before migration takes hold. The magnitude
of the spillovers modestly depends on the instrument used (with IV II
producing larger estimates in all models except for 6-year differences),
though the results are generally robust to instrument choice (except in
the six-year differenced metro results; which we mostly omit from the
discussion due to the weak instruments, except when we discuss the
Lewbel (2012) instruments and OLS below).
The three-year first difference nonmetropolitan results in columns
[3] and [4] indicate that an expanding oil and gas sector has about the
same multiplier effects as the one-year models (ignoring the relatively
small three-year coefficient in column [3]). However, the six-year first
difference models suggest the nonmetro energy multiplier is above
three, implying significant spillovers that might relate to a lag between
initial drilling and construction build-up in infrastructure such as pipelines. Given that migration takes time to respond, nonmetro labour markets may have larger responses because of in-migrants filling labour
demand that create additional spillovers. The ten-year multiplier estimates with instrument set IV I suggest (insignificant) crowding out,
20
A value of 10 is often used as a rule-of-thumb cutoff value for the first stage F-statistic,
which (if N10) indicates a strong instrument(s) (Stock et al., 2002). First-stage results are
available on request.
21
Our main interest, however, is the IV nonmetro results.
22
It is particularly interesting that the metropolitan models appear the most problematic in terms of the over-identification tests as oil and gas employment is a small proportion
of total metro employment. Given the Durbin endogeneity test results, one may want to
rely on the OLS metro results below.
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
87
Table 1
2SLS estimation results (DV is the change in total employment growth rates between periods).
1-year differences
IV I [1]
Nonmetropolitan areas
DiffEnVar
1.261***
(0.180)
DiffDShock
1.517***
(0.245)
Mining85
0.007**
(0.003)
IV F-stat
317.9
Durbin P
0.212
Overid P
0.139
a
0.180
Wald P
# Obs
21,846
Metropolitan areas
DiffEnVar
DiffDShock
Mining85
IV F-stat
Durbin P
Overid P
Wald Pa
# Obs
0.635
(0.448)
1.151***
(0.114)
0.005
(0.005)
175.3
0.720
0.000
0.295
11,429
3-year differences
6-year differences
10-year differences
IV II [2]
IV I [3]
IV II [4]
IV I [5]
IV II [6]
IV I [7]
IV II [8]
1.310***
(0.175)
1.518***
(0.245)
0.007**
(0.003)
2,045.0
0.150
0.000
0.313
0.642**
(0.294)
2.063***
(0.328)
0.147***
(0.032)
27.4
0.002
0.001
0.001
5,958
1.216***
(0.349)
2.031***
(0.311)
0.125***
(0.028)
29.1
0.255
0.000
0.085
3.713***
(0.866)
1.536***
(0.187)
0.320***
(0.097)
16.1
0.046
0.023
0.023
1,986
3.037***
(0.366)
1.558***
(0.178)
0.334***
(0.085)
4.2
0.012
0.049
0.000
0.290
(0.570)
1.600***
(0.142)
0.230
(0.284)
13.1
0.033
0.973
0.017
1,986
1.734***
(0.408)
1.718***
(0.138)
−0.356*
(0.201)
2.4
0.407
0.000
0.969
0.701
(0.441)
1.150***
(0.114)
0.005
(0.005)
1,208.0
0.838
0.000
0.353
0.402
(0.767)
1.389***
(0.388)
0.144***
(0.053)
19.3
0.420
0.005
0.278
3,117
0.748
(0.720)
1.364***
(0.386)
0.134**
(0.052)
76.2
0.684
0.000
0.473
−3.040
(20.668)
2.400***
(0.380)
0.516
(0.405)
0.1
0.708
0.062
0.794
1,039
10.594
(15.988)
2.305***
(0.372)
0.758**
(0.369)
0.4
0.319
0.161
0.606
0.851
(0.859)
1.940***
(0.251)
−0.076
(0.370)
14.9
0.605
0.424
0.231
1,039
0.608
(0.879)
1.936***
(0.250)
−0.002
(0.387)
11.6
0.750
0.268
0.154
***, **, *, significant at 0.01, 0.05, and 0.1, respectively; standard errors clustered at BEA area level in parentheses; all models include log of population in 1980, 1990 employment shares in
manufacturing and agriculture, 1990 shares of adult population with only high school diploma and with a BA degree or higher, as well as time period dummies. aP-value from the Wald test
of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
whereas instrument set II indicates positive spillovers on net with a
multiplier of 1.73.
The overall time path of the expanding energy sector effects appears
to be the following: modest short-term positive net spillovers in the
one to three-year time frame, then rising to a large multiplier of 3 in
the course of six years (though the multiplier is smaller in sensitivity
analysis), and a decline with small positive spillovers after ten years.
As noted earlier, because the time period does not include a bust, we
cannot assess the longer-run aspects over the boom–bust cycle.
To assess whether energy shocks have similar impacts as equally
sized shocks across the rest of the economy, we compare the energy
(DiffEnVar) and industry mix (DiffDShock) coefficients. In terms of coefficient magnitude, the industry mix coefficient tends to be larger than
the energy coefficient in all models except for the nonmetro six-year
first differences. Table 1 and all other tables that show IV estimation results report Wald test of the equality between DiffEnVar and DiffDShock
(Wald P for a test of H0: βDiffEnVar = βDiffDShock). They suggest that in
most nonmetropolitan cases, the difference in coefficients is statistically
significant, while this is not the case in the metropolitan results. Yet, because this pattern almost always prevails across all equations and across
all subsectors, in the sensitivity analysis section, we use a seemingly unrelated regression (SUR) model to test the hypothesis of coefficient
equality in a system approach. In virtually all cases, we find that the
null can be rejected at the 0.001% level, providing confidence in these
patterns.
The smaller energy shock coefficient suggests that in the shorter
1- to 3-year range and in the longer 10-year first differences, the energy shocks have smaller net-positive spillovers to other local industries
compared to the average shocks elsewhere in the economy. This may
happen due to the “shallow” nature of the input–output energy supply
chain in rural areas where even with the relatively high wages in the energy sector together with lease and royalty payments to landowners,
much of the inputs are imported from elsewhere. Alternatively, the
workers may reside outside of the rural communities and commute
in, creating income leakages. For six-year differences, the pattern is reversed with the energy sector having larger net positive spillovers,
suggesting that the short-term construction of infrastructure is important for job growth, though this effect is temporary.
To provide some perspective, the average nonmetropolitan county
gained 0.08 percent net total employment growth from energy-sector
expansion during the 2001–2013 period. Using the one-year difference
coefficients in Table 1, this translates into about 6.5 net jobs per year
created by the energy boom when evaluated at the median nonmetro
county employment of 6,569, equalling to 0.1 percent increase in total
annual job growth. Focusing on the nonmetro counties in the top 20
percent of energy sector expansion, total job growth is expected to be
0.34 higher as a direct result of energy employment growth, which
translates into about 25.9 net jobs created on an annual basis when
evaluated at the corresponding median nonmetro county employment
of 5,876. In sum, given that the average representative shock tends to
have larger employment effects and likely leads to a more diversified
economy, focusing more broadly on other sectors to promote growth
seems to be the best way forward for rural communities desiring economic development.
The 1985 mining share variable accounts for long-term historical effects of having the industry in a location. A priori, the expected sign is
ambiguous. It could have positive contemporaneous effects as the associated infrastructure is already in place, allowing energy development
to occur more easily. However, if the associated infrastructure is at
least partially in place, there may be less need for construction and support workers. Table 1 shows that the 1985 mining share coefficient is insignificant in the ten-year difference models, but is significantly positive
and increases over time in other models. This can be interpreted as limited evidence of positive legacy mining effects during booms.
In models for metropolitan counties, the oil and gas multiplier is statistically insignificant, although in some sectors, as demonstrated
below, there are positive spillovers. As noted by Munasib and Rickman
(2015), the impact of changes in energy employment is likely to be
less pronounced in urban areas due to the sheer size of their economies.
Unlike the insignificant energy coefficient, the industry mix coefficient
suggests positive spillovers from general shocks on average regardless
of the differencing span.
88
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
Table 2
2SLS estimation results (DV is the change in total population growth between periods).
1-year differences
IV I [1]
Nonmetropolitan areas
DiffEnVar
−0.288***
(0.074)
DiffDShock
0.065***
(0.015)
Mining85
0.005***
(0.002)
IV F-stat
317.9
Durbin P
0.000
Overid P
0.000
0.000
Wald Pa
# Obs
21,846
Metropolitan areas
DiffEnVar
DiffDShock
Mining85
IV F-stat
Durbin P
Overid P
Wald Pa
# Obs
−0.608***
(0.230)
0.004
(0.026)
0.006***
(0.002)
175.3
0.004
0.824
0.013
11,429
3-year differences
6-year differences
10-year differences
IV II [2]
IV I [3]
IV II [4]
IV I [5]
IV II [6]
IV I [7]
IV II [8]
−0.229***
(0.071)
0.065***
(0.015)
0.005***
(0.002)
2,045.0
0.003
0.000
0.000
0.184**
(0.074)
0.249***
(0.039)
0.011**
(0.004)
27.4
0.377
0.008
0.384
5,958
0.230***
(0.075)
0.246***
(0.039)
0.010**
(0.004)
29.1
0.083
0.009
0.841
0.513***
(0.077)
0.038**
(0.020)
0.015
(0.021)
16.1
0.000
0.931
0.000
1,986
0.252***
(0.065)
0.047***
(0.018)
0.020
(0.016)
4.2
0.052
0.014
0.001
0.159***
(0.059)
0.043***
(0.013)
−0.004
(0.026)
13.1
0.412
0.291
0.054
1,986
0.199***
(0.043)
0.047***
(0.014)
−0.020
(0.022)
2.4
0.120
0.020
0.001
−0.690**
(0.273)
0.004
(0.026)
0.006***
(0.002)
1,208.0
0.006
0.161
0.017
−0.140
(0.104)
0.208**
(0.081)
0.031***
(0.006)
19.3
0.076
0.558
0.006
3,117
−0.110
(0.094)
0.206**
(0.081)
0.030***
(0.006)
76.2
0.118
0.001
0.008
1.998
(4.619)
0.081
(0.057)
0.079
(0.096)
0.1
0.182
0.963
0.679
1,039
−0.000
(1.209)
0.095*
(0.055)
0.044*
(0.026)
0.4
0.956
0.388
0.938
0.089
(0.088)
0.078***
(0.022)
0.008
(0.033)
14.9
0.950
0.884
0.902
1,039
0.070
(0.090)
0.078***
(0.022)
0.014
(0.033)
11.6
0.869
0.054
0.929
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
4.2. Results for population
Table 2 shows the results where employment growth is replaced
with population growth as the dependent variable. Considering population growth allows us to assess whether the net job creation primarily
goes to the original residents of the county or to outsiders such as
new migrants (which do not include “fly-ins” who reside elsewhere).
If the newly created jobs are filled by migrants, the potential of the
local employment growth to improve local opportunities for those
who are underemployed or out of the labour force is undermined.23
Before describing the specific results, we note that the diagnostic
statistics are similar to total employment with (for instance) consistently strong instruments except for IV II in 6- and 10-year nonmetro
models and the 6-year metro models. The one-year difference nonmetropolitan results suggest that there is a small net out-migration in response to energy development, perhaps by those who do not desire
some of the negative externalities associated with drilling.24 Over
the three-, six-, and ten-year differences, however, the population
“multiplier” starts at about 0.2, peaking at perhaps as high as 0.5 at six
years, before falling to just under 0.2 after 10 years (all statistically
significant). Nonetheless, these very small population responses compared to the employment responses in Table 1 suggest that most of
the new net jobs created in nonmetro counties go either to original residents, commuters from nearby counties, or “fly-ins” who might temporarily reside in the county. The lack of sizable population increase as a
result of new energy development limits the positive agglomeration
23
Population changes have the advantage of being the most accurate measure compared
to estimates of net domestic migration or international migration. It is also the most comprehensive measure.
24
One possible example of such net out-migration (or at least not strong in-migration) is
Bradford County, PA, which has been one of the most exposed Marcellus shale play
counties in terms of having over 1,000 wells drilled. Yet, between 2006 (before drilling began in earnest) and 2014, Bradford county population declined from 62,345 to 61,784,
while US population rose 6.9%. Of course, this suggests that boomtowns like Williston,
North Dakota are far outliers, in which Williams County (Williston is the county seat)
went from 20,122 to 29,595 between 2006 and 2013 using the US Census Bureau data.
effects that could reduce the population outflow during a bust and permanently help in the long run.
Compared to a representative shock in the economy, the population
inflow response to energy shocks is smaller in the three-year period and
slightly larger in the six- and ten-year periods, though both shocks lead
to small population change. This is consistent with a significant decline
in migration response to economic shocks after the year 2000 reported
by Partridge et al. (2012). It follows that the long-term agglomeration
effects hypothesized by Michaels (2011) are much less likely for modern boomtowns.
For metro counties the one-year first difference results suggest
about a −0.6 population multiplier in response to energy shocks, consistent with short-term out-migration as a result of oil and gas development. These multipliers in models with wider differencing spans are
statistically insignificant. Thus, energy booms are not associated with
large metro population expansions. This is expected because a rapidly
growing energy sector employs a tiny fraction of the population in a
vast majority of metro counties. By contrast, the representative
economy-wide shock is associated with modest population growth
that is statistically significant in three-year and ten-year differenced
models.
4.3. Results for selected sectors
Table 3 reports 2SLS results for the traded goods sector defined here
as agriculture, mining (except NAICS2111 and NAICS2131), and selected manufacturing industries (Appendix A provides more details).
Before turning to the results, note that the first-stage F-statistics
imply strong instruments (except for the IV II in 6- and 10-year differences models in the nonmetro sample and 6-year metropolitan sample;
which we do not discuss) and the over-identification tests generally indicate well-identified models. The Dutch Disease arguments suggest
that energy sector growth bids up wages, reducing growth in the traded
goods sector as their costs increase. At the same time, local demand
from energy industries could stimulate manufacturing, offsetting the
negative impact of increased production costs. The oil and gas coefficient is almost always negative in the nonmetro sample. It is significant
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
89
Table 3
2SLS estimation results (DV is the change in tradable industries employment growth rates between periods).
1-year differences
IV I [1]
Nonmetropolitan areas
DiffEnVar
−0.112
(0.141)
DiffDShock
0.966***
(0.224)
Mining85
0.001
(0.002)
IV F-stat
317.9
Durbin P
0.437
Overid P
0.255
a
0.000
Wald P
# Obs
21,846
Metropolitan areas
DiffEnVar
DiffDShock
Mining85
IV F-stat
Durbin P
Overid P
Wald Pa
# Obs
−0.256
(0.203)
0.522***
(0.059)
−0.001
(0.003)
175.3
0.468
0.546
0.000
11,429
3-year differences
6-year differences
10-year differences
IV II [2]
IV I [3]
IV II [4]
IV I [5]
IV II [6]
IV I [7]
IV II [8]
−0.070
(0.128)
0.967***
(0.224)
0.001
(0.002)
2,045.0
0.599
0.232
0.000
−0.253***
(0.087)
0.838***
(0.139)
0.040**
(0.016)
27.4
0.007
0.019
0.000
5,958
−0.128*
(0.071)
0.831***
(0.137)
0.035**
(0.015)
29.1
0.081
0.000
0.000
−0.067
(0.173)
0.456***
(0.086)
0.029
(0.037)
16.1
0.840
0.586
0.014
1,986
0.097
(0.097)
0.451***
(0.085)
0.026
(0.036)
4.2
0.017
0.903
0.011
−1.323***
(0.389)
0.509***
(0.077)
0.386*
(0.228)
13.1
0.000
0.653
0.000
1,986
−0.529**
(0.209)
0.574***
(0.073)
0.063
(0.118)
2.4
0.133
0.000
0.000
−0.247
(0.195)
0.522***
(0.059)
−0.001
(0.003)
1,208.0
0.482
0.124
0.000
0.126
(0.253)
0.227**
(0.105)
−0.006
(0.028)
19.3
0.493
0.379
0.672
0.140
(0.246)
0.226**
(0.105)
−0.006
(0.028)
76.2
0.435
0.330
0.709
3,117
−2.320
(6.933)
0.556**
(0.238)
−0.106
(0.166)
0.1
0.570
0.504
0.670
4.775
(9.968)
0.507**
(0.237)
0.020
(0.206)
0.4
0.376
0.520
0.682
1,039
−0.916
(0.685)
0.175**
(0.074)
0.160
(0.325)
14.9
0.291
0.405
0.019
−0.948
(0.665)
0.174**
(0.074)
0.170
(0.320)
11.6
0.260
0.765
0.012
1,039
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
at the 0.05 level in three cases including both 10-year models. Yet, the
industry mix coefficient indicates that the average industry shock has
a consistent positive and statistically significant effect on traded goods
job growth. In the metro results, oil and gas coefficients are insignificant,
whereas the industry mix term is consistently positive and significant.25
Especially when compared to larger positive net response to the average
shock elsewhere in the economy, the results suggest that oil and gas development crowds out jobs in tradable industries in the longer-run,
weakly consistent with Dutch Disease effects.
Table 4 reports the non-traded goods sector results. The diagnostic
statistics show that the nonmetro and metro models are better identified for the 10-year differencing, although only in the shorter period
nonmetro models is there evidence of endogeneity.
The results for nonmetropolitan counties indicate that oil and gas
have a statistically significant stimulating effect (except for the 3-year
differenced models) on employment in non-tradable industries. This
positive response to energy sector growth is smaller than the average
industry mix effect in the 1- and 3-year models, but it exceeds the average effect in the 6- and 10-year models. Combining the results for traded
and non-traded goods, we conclude that a possible reason for observing
an overall multiplier larger than the energy multiplier on average is that
in the shorter periods, the effects are observable in both traded and nontraded goods, whereas in longer periods, it appears to be primarily due
to the energy sector crowding out other traded goods. The Wald test
suggests that the spillover effects of energy and overall economic shocks
are not statistically different in the 1-year differenced models; they are
statistically different in the 3- and 6-year differenced models with
mixed evidence for the 10-year differenced models. Overall, in contrast
to the evidence for traded goods sector, oil and gas development in nonmetro counties tends to have longer-term positive effects on employment in non-tradable industries.
25
Appendix Tables E.1 and E.2, respectively, present estimation results for agriculture
(crowding out after 10 years) and manufacturing (weak evidence of crowding out after
three years).
For metro counties, the models show no impact on non-traded
goods employment from the expansion in the oil and gas industry in
the short and medium run, although in the MSAs that are “outliers,”
like Houston and Tulsa, outsized stimulating effects on non-traded employment are likely present. After 10 years, on the contrary, oil and gas
industry growth stimulates total employment in the non-tradable sector. The average shock from other industries outside energy consistently
boosts non-traded goods employment as well.
Construction is arguably a non-tradable sector that benefits the most
from oil and gas development. Table 5 reports corresponding results for
this industry. The nonmetro results suggest a consistent positive response to energy sector growth that exceeds that of the average industry shock with the exception of the three-year models, although in
longer durations, the Wald test cannot reject the null that the two
shocks have equal-sized effects. The relative magnitude of the effects
in Tables 4 and 5 indicate that the positive effect of energy sector growth
on non-traded goods sector is mostly through construction. The metro
results indicate positive effect of energy on construction employment
after 1 and 10 years with no statistically significant pattern in the
other periods.
Appendix Tables E.3 through E.10 report the results for non-traded
goods sector by industry. Focusing on some noteworthy results, Appendix Table E.3 shows that transportation and warehousing in nonmetro
counties consistently enjoy positive employment spillovers from energy
sector growth. The large needs for transporting products to well sites,
especially water used for shale development, likely explain this positive
relationship. The retail and wholesale trade industries (Tables E.4 and
E.5) in rural areas enjoy positive spillovers with a delay of 3 to 6 years,
consistent with a lagged agglomeration effect. Given that many of the
oil and gas workers live in temporary housing and are likely eat out regularly, it is unsurprising that accommodation employment, as well as
real estate, are particularly stimulated in nonmetropolitan areas by energy sector development (Tables E.6 and E.7).
Table 6 focuses directly on the upstream industries that most intensively supply the oil and gas industry, following the classification described in Appendix Table A.2. With the exception of the ten-year
90
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
Table 4
2SLS estimation results (DV is the change in non-tradable industries employment growth rates between periods).
1-year differences
IV I
Nonmetropolitan areas
DiffEnVar
0.373***
(0.100)
DiffDShock
0.551***
(0.083)
Mining85
0.005**
(0.003)
IV F-stat
317.9
Durbin P
0.007
Overid P
0.048
a
0.120
Wald P
# Obs
21,846
Metropolitan areas
DiffEnVar
DiffDShock
Mining85
IV F-stat
Durbin P
Overid P
Wald Pa
# Obs
−0.109
(0.369)
0.629***
(0.079)
0.006
(0.004)
175.3
0.984
0.000
0.062
3-year differences
6-year differences
10-year differences
IV II
IV I
IV II
IV I
IV II
IV I
IV II
0.380***
(0.102)
0.551***
(0.083)
0.005**
(0.003)
2,045.0
0.017
0.002
0.150
−0.105
(0.261)
1.225***
(0.289)
0.107***
(0.024)
27.4
0.009
0.001
0.000
5,958
0.345
(0.313)
1.200***
(0.278)
0.090***
(0.020)
29.1
0.505
0.000
0.038
2.779***
(0.904)
1.080***
(0.169)
0.291***
(0.094)
16.1
0.048
0.009
0.088
1,986
1.940***
(0.319)
1.106***
(0.158)
0.308***
(0.076)
4.2
0.024
0.004
0.014
1.573***
(0.362)
1.046***
(0.106)
−0.208
(0.178)
13.1
0.100
0.326
0.120
1,986
1.575***
(0.312)
1.046***
(0.104)
−0.209
(0.159)
2.4
0.175
0.142
0.075
−0.052
(0.363)
0.628***
(0.079)
0.006
(0.004)
1,208.0
0.894
0.000
0.080
11,429
−0.724
(0.708)
1.162***
(0.357)
0.150***
(0.035)
19.3
0.273
0.010
0.022
−0.392
(0.666)
1.138***
(0.355)
0.140***
(0.033)
76.2
0.492
0.000
0.048
3,117
−1.719
(14.485)
1.844***
(0.224)
0.621**
(0.262)
0.1
0.803
0.018
0.807
4.819
(8.981)
1.798***
(0.220)
0.738***
(0.209)
0.4
0.605
0.062
0.738
1,039
2.002**
(0.968)
1.604***
(0.191)
−0.232
(0.308)
14.9
0.119
0.664
0.686
1,039
1.911**
(0.916)
1.602***
(0.191)
−0.204
(0.293)
11.6
0.130
0.732
0.742
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
nonmetro models, the response to oil and gas shocks is statistically insignificant and less than the corresponding response to the average
shock (which is consistently statistically significant). Oil and gas development appears to be ineffective at stimulating local employment in
supply-chain industries compared to the average industry. Even at ten
years, the energy multiplier for upstream industries only ranges from
0.140 to 0.223, or for every 100 jobs created in the local energy industry,
only about 14 to 22.3 jobs are expected to be created in upstream supply
industries during this period. It may be capturing the fact that upstream
oil and gas industry suppliers often locate outside of the local area.
5. Sensitivity analysis
To assess the robustness of our results, we perform a series of sensitivity analyses using different approaches and different samples.
Because most Durbin p-values for endogeneity are insignificant in
Table 5
2SLS estimation results (DV is the change in construction employment growth rates between periods).
1-year differences
IV I
Nonmetropolitan areas
DiffEnVar
0.289***
(0.061)
DiffDShock
0.094***
(0.026)
Mining85
0.004**
(0.002)
IV F-stat
317.9
Durbin P
0.000
Overid P
0.009
0.000
Wald Pa
# Obs
21,846
Metropolitan areas
DiffEnVar
DiffDShock
Mining85
IV F-stat
Durbin P
Overid P
Wald Pa
# Obs
0.428**
(0.172)
0.124***
(0.030)
0.000
(0.002)
175.3
0.001
0.000
0.081
3-year differences
6-year differences
10-year differences
IV II
IV I
IV II
IV I
IV II
IV I
IV II
0.274***
(0.054)
0.094***
(0.026)
0.004**
(0.002)
2,045.0
0.000
0.018
0.000
0.098
(0.091)
0.484***
(0.116)
0.041***
(0.010)
27.4
0.399
0.016
0.002
5,958
0.183*
(0.104)
0.479***
(0.115)
0.038***
(0.010)
29.1
0.716
0.001
0.025
0.869***
(0.233)
0.529***
(0.064)
0.124***
(0.033)
16.1
0.015
0.129
0.209
1,986
0.509***
(0.113)
0.540***
(0.064)
0.132***
(0.028)
4.2
0.056
0.070
0.788
0.376***
(0.104)
0.291***
(0.035)
−0.038
(0.048)
13.1
0.319
0.172
0.424
1,986
0.443***
(0.088)
0.296***
(0.036)
−0.066
(0.041)
2.4
0.147
0.446
0.111
0.430**
(0.168)
0.124***
(0.030)
0.000
(0.002)
1,208.0
0.000
0.000
0.073
11,429
−0.008
(0.253)
0.592***
(0.151)
0.045***
(0.011)
19.3
0.380
0.024
0.062
0.082
(0.248)
0.586***
(0.150)
0.042***
(0.011)
76.2
0.601
0.000
0.112
3,117
0.518
(3.811)
0.663***
(0.072)
0.191***
(0.073)
0.1
0.943
0.000
0.970
1.830
(3.592)
0.654***
(0.068)
0.215***
(0.083)
0.4
0.601
0.019
0.744
1,039
0.568**
(0.265)
0.380***
(0.040)
−0.063
(0.084)
14.9
0.085
0.702
0.474
0.576**
(0.252)
0.380***
(0.040)
−0.066
(0.080)
11.6
0.070
0.409
0.434
1,039
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
91
Table 6
2SLS estimation results (DV is the change in upstream to energy sector employment growth rates between periods).
1-year differences
IV I
Nonmetropolitan areas
DiffEnVar
−0.043
(0.045)
DiffDShock
0.137***
(0.026)
Mining85
0.001**
(0.000)
IV F-stat
317.9
Durbin P
0.907
Overid P
0.010
0.002
Wald Pa
# Obs
21,846
Metropolitan areas
DiffEnVar
DiffDShock
Mining85
IV F-stat
Durbin P
Overid P
Wald Pa
# Obs
0.137
(0.124)
0.113***
(0.028)
0.000
(0.001)
175.3
0.468
0.537
0.851
3-year differences
6-year differences
10-year differences
IV II
IV I
IV II
IV I
IV II
IV I
IV II
−0.038
(0.040)
0.137***
(0.026)
0.001**
(0.000)
2,045.0
0.762
0.062
0.001
−0.016
(0.050)
0.208***
(0.059)
0.010***
(0.003)
27.4
0.632
0.219
0.001
5,958
0.009
(0.042)
0.206***
(0.059)
0.009***
(0.003)
29.1
0.935
0.121
0.001
0.091
(0.082)
0.130***
(0.033)
0.026**
(0.013)
16.1
0.714
0.158
0.667
1,986
0.032
(0.035)
0.132***
(0.033)
0.027**
(0.013)
4.2
0.320
0.396
0.023
0.223***
(0.062)
0.233***
(0.024)
−0.045
(0.030)
13.1
0.010
0.840
0.869
1,986
0.140***
(0.030)
0.227***
(0.024)
−0.011
(0.019)
2.4
0.168
0.355
0.026
0.124
(0.122)
0.113***
(0.028)
0.000
(0.001)
1,208.0
0.522
0.482
0.928
11,429
0.104
(0.165)
0.139
(0.122)
0.018***
(0.006)
19.3
0.568
0.115
0.880
0.145
(0.168)
0.136
(0.122)
0.016***
(0.006)
76.2
0.377
0.037
0.969
3,117
−3.244
(7.457)
0.191***
(0.074)
0.016
(0.145)
0.1
0.156
0.340
0.647
−2.366
(3.805)
0.185***
(0.058)
0.031
(0.078)
0.4
0.244
0.595
0.506
1,039
0.103
(0.186)
0.369***
(0.030)
−0.003
(0.052)
14.9
0.564
0.225
0.142
0.092
(0.176)
0.369***
(0.030)
0.001
(0.050)
11.6
0.584
0.625
0.107
1,039
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
Table 1, perhaps the most natural starting point is comparing our findings to OLS. Table 7 presents the OLS results for total employment. In
nonmetro areas, the magnitude of the effects follows the same pattern
as the one reported in Table 1 with a continuous increase between
one and six years followed by a decrease in the 10-year analysis. The average shock generally has a larger estimated multiplier than the one for
the energy sector. One result that stands out is that the nonmetro sixyear OLS energy multiplier is 2.23, which is considerably less than the
IV estimate.
Recall that the Durbin test suggested that endogeneity was not a severe problem in the metro models, implying that the OLS results may be
more germane. In this case, unlike the IV results, the DiffEnVar coefficient is statistically significant in all periods except for the 10-year
differenced model. It reveals net crowding out in the one-year and positive net spillovers in the 3- and 6-year models. Overall, while there are
differences in the precision of the estimates (with OLS tending to have
smaller standard errors), the average energy and overall DiffDShock coefficients are very close in both the IV and OLS models, which at the very
least suggests that measurement error is not dominating the results.26
To further test for the equality of the DiffEnVar and the DiffDShock coefficients, we run several SUR models for the nonmetro and metro samples that jointly estimate the total employment models with the
sectoral models (not shown).27 By jointly modelling total employment
and sectoral employment, we can take advantage of any correlation in
the residuals to improve the efficiency in the estimates, though mostly
26
For example, for nonmetropolitan total employment, across the years and different IV
specifications, the average energy multiplier is 1.65 versus 1.57 for OLS, while the average
IV DiffDShock multiplier equals 1.69, with the average for OLS is 1.70. For metropolitan total employment, the average energy IV multiplier is 0.66, while the corresponding OLS average equals 0.68 (not including the 6-year estimates in which the IV instruments are
weak). Similarly, for the average DiffDShock multiplier, the average metropolitan multiplier is 1.70 in both the IV models and the OLS models.
27
The specific combinations tried for the metro and nonmetro samples are: total employment, traded, and upstream industries; total employment, nontraded, and upstream
industries; total employment and all of the 14 individual sectors. We can’t estimate the total employment and traded and nontraded sectors together because they are perfectly
collinear.
we want to benefit from having the extra power from more observations for testing the (joint) null hypothesis. In virtually every differencing case in both the nonmetro and metro samples, we find that null
hypothesis is usually rejected at the .001% level and in the other cases,
it can always be rejected at the 5% level, further providing support for
our general claim that average shocks elsewhere in the economy produce larger spillovers than equal-sized energy shocks.
OLS estimation results for tradable and non-tradable sectors are presented in Tables D.1 and D.2, respectively. In line with the results reported in Table 3, an expanding energy sector seems to crowd out tradable
sector employment, although the OLS estimates reported in Appendix
D tend to be weaker in statistical significance and smaller in magnitude.
The OLS and IV results for non-traded goods, however, are fairly
consistent.
Our next step is to check whether any spatial dependence affects the
results by first examining the Moran’s I statistic. In general, Moran’s I
statistic points to the presence of spatial autocorrelation in some sectors
and in some years with the evidence of spatial autocorrelation becoming stronger with wider differencing spans and in later periods. To see
if this affects our main conclusions, we include spatial lags of the main
explanatory variables, DiffEnVar and DiffDShock in the main specification (the results are available on request). We consider spillovers from
all contingent counties. The spatial lags of the explanatory variables calculated using corresponding values for all contingent counties are
scaled by the county’s total employment. In this case, estimation coefficients on the spatially lagged variables can be interpreted as multipliers
and directly compared to the size of the respective own-county coefficients. The instruments for WDiffEnVar are defined in the same manner
as the own county DiffEnVar instruments but using the values from the
contingent counties. We find that the IV I (WIV I instrument) set with
fewer instruments provides much stronger identification. For this
case, the magnitude of the coefficients when scaled to the same owncounty total employment benchmark as the main explanatory variables,
however, is small, suggesting very minor multipliers and spillovers from
surrounding counties. For the total employment equation in the nonmetropolitan sample, coefficients on the spatially lagged energy variable range from 0.0105 in the 1-year differenced models to as high as
92
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
Table 7
OLS results for full models (DV is the change in total employment growth rates between periods).
1-year differences
DiffEnVar
DiffDShock
Mining85
R2
Wald Pa
Clusters
Observations
3-year differences
6-year differences
10-year differences
Nonmetro
Metro
Nonmetro
Metro
Nonmetro
Metro
Nonmetro
Metro
1.031*** [0.073]
1.515*** [0.245]
0.007** [0.003]
0.087
0.054
174
21,846
0.790*** [0.193]
1.150*** [0.114]
0.005 [0.004]
0.176
0.113
161
11,429
1.482*** [0.197]
2.016*** [0.306]
0.115*** [0.025]
0.305
0.145
174
5,958
1.048*** [0.188]
1.343*** [0.384]
0.125** [0.049]
0.510
0.492
161
3,117
2.228*** [0.303]
1.583*** [0.185]
0.350*** [0.080]
0.351
0.055
174
1,986
1.789*** [0.445]
2.367*** [0.355]
0.602*** [0.174]
0.258
0.326
161
1,039
1.549*** [0.348]
1.703*** [0.135]
−0.281* [0.166]
0.325
0.668
174
1,986
0.211 [0.890]
1.930*** [0.260]
0.118 [0.471]
0.316
0.032
161
1,039
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
0.0339 in the 3-year differenced models (both are significant at the 1%
level).28 The corresponding metro energy results are of similar magnitude but they are not as precisely estimated. Perhaps more importantly,
the key own-county energy shock and average employment shock results remain virtually unchanged, consistent with our expectation that
our identification strategy accounts for any omitted variables such as
spatial lags.
To further examine our identification, we now consider the GMM instrumental variable estimation proposed by Lewbel (2012) using Stata’s
IVREG2H procedure. Intuitively in Lewbel’s method, unobserved
heterogeneity is reflected in heteroskedasticity of the error terms.
This heteroskedasticity is used to form alternative instruments
(Carmignani, 2014). When using these constructed instruments in combination with the instruments used in this study, the results follow the
same basic patterns as before, though the estimates produced by combined IV I and IV II are closer to each other (results not shown for brevity). The Lewbel nonmetro energy multipliers tend to be a little smaller
and the metro multipliers tend to be a little larger than for the reported
results (their significance is typically improved as well) with the instruments also being stronger than before. Yet, one difference is in the sixyear difference models in which the nonmetro multipliers are just
over 2.2 and the metro energy multipliers are just over 1.8 (consistent
with the OLS results). Thus, while these results support the robustness
of the IV results (including finding that the average economy-wide multiplier exceeds the energy multiplier), they also tend to suggest the OLS
six-year results are more credible.
We then test whether our instruments may have redundancies
given the time period interactions that produce weak instruments. We
first note that other indicators of the weak instruments such as the
Kleibergen–Paap Wald rk F statistics typically suggest strong firststage instruments in our base total employment model (values N10),
consistent with the standard F-statistic, though the values were sometimes smaller. Thus, to further assess our identification, we also consider
fewer instruments by dropping some of the “redundant” instruments in
the first stage by using the IVREG2 redundant option in the one- and
three-year differencing model. This procedure omits instruments that
are statistically redundant and add little to the identification (the results
are omitted for brevity but are available from the authors).
To summarize, the results of this exercise suggest that the average
economy-wide multiplier and the energy standard errors are slightly
smaller in this case but the results are very similar. Two exceptions
are that the three-year nonmetro differences now suggest a smaller energy sector multiplier (B = 0.54, se = .186) and the one-year metro energy multiplier appears to be larger (B = 1.08, se = .396). In addition,
the over-identification tests suggest these models are better identified.
Regarding the strength of the instruments in the first stage, the results
are mixed with some cases being stronger, but many others are weaker
than the base IV model. Yet, the Wooldridge test suggests that the energy variable is exogenous, with the exception of the three-year nonmetro
28
A full set of estimation results for the equations that include spatially lagged explanatory variables and for the Moran’s I testing is available upon request.
models. The overall story, however, is quite similar regardless of using
IV or OLS, which is reassuring in that neither significant changes in the
number of instruments themselves nor changes in the strength of the
instruments have any tangible effect on our main conclusions.
As the next step, we examine a parsimonious model in which we
omit all of the variables except the energy, industry mix, mining share
in 1985, and the time period dummies (see Appendix Table F.1).
These results are quite similar to the base IV results in Table 1. Next,
to test for a nonlinearity in oil and gas multiplier (maybe due to
crowding out or agglomeration), we added DiffEnVar squared to the
base IV model (see Table F.2 in the Appendix). There is weak evidence
of nonlinearity in only a few cases but overall it seems that a linear approximation is appropriate, and thus we do not consider this model further. As a final sensitivity test, we take the base IV model and omit the
1985 mining share and industry mix variable one-by-one to see if it affects the results. Like in the previous case, the estimates stay very close
to those reported in Table 1 (Tables F.3–F.4).
To assess the robustness of our results in various subsamples, we
first estimate the model excluding some of the most “powerful” observations and outliers in terms of oil and gas endowment and production.
For the nonmetro sample, we first exclude all counties that are above
the Bakken play. These results are displayed in Table 8; they indicate
that dropping the Bakken modestly reduces the magnitude of the multipliers. Otherwise, the results reported in Table 1 are robust.29
Dropping large metropolitan energy centres (Dallas, Houston, Oklahoma
City, and Tulsa) from the metro sample appreciably changes the significance of the energy coefficient in only one case, in which crowding out
as a result of energy sector expansion using IV II in the 1-year analysis
becomes weakly significant.
To test for the potential differences between oil and gas employment
multipliers we create a subsample of predominantly oil-producing
counties, those sitting above the Bakken, Eagle Ford, and Permian
plays, and a subsample of predominantly gas-producing counties,
those above the Marcellus, Haynesville, and Niobrara plays.30 The number of metropolitan counties in the predominantly oil- and predominantly gas-producing subsamples is too small for a credible inference.
Likewise, because the number of nonmetro counties in both samples
is larger than in the metro sample, but still small (63 for oil producing;
101 for gas producing counties), the results should be cautiously
interpreted. We cannot use IV in these models because the geological
and historical conditions have almost no variation in these subsamples
due to large oil or gas endowments, producing very weak instruments
measured by the F-stat in the first-stage. Table 9, therefore, reports
29
We also estimated the model only for nonmetro counties that lie above Bakken, Eagle
Ford, and Barnett shale plays. The strength of the instruments in all of these models is below 10, so we do not report the results.
30
The dichotomy between oil and gas producing is not perfect as all of these main shale
plays produce some oil or some gas. We made our determination based on relative production data from the US EIA (see http://www.eia.gov/petroleum/drilling/#tabssummary-2). For a US EIA map of these shale plays, see www.eia.gov/oil_gas/rpd/shale_
gas.pdf.
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
93
Table 8
2SLS estimation results for alternative samples (DV is the change in total employment growth rates between periods).
1-year differences
IV I [1]
Nonmetropolitan areas, no Bakken
DiffEnVar
1.190***
(0.176)
DiffDShock
1.516***
(0.252)
Mining85
0.007**
(0.003)
IV F-stat
421.6
Durbin P
0.353
Overid P
0.211
a
0.073
Wald P
# Obs
21,329
3-year differences
6-year differences
10-year differences
IV II [2]
IV I [3]
IV II [4]
IV I [5]
IV II [6]
IV I [7]
IV II [8]
1.288***
(0.179)
1.517***
(0.252)
0.006**
(0.003)
2,508.0
0.138
0.001
0.248
0.493*
(0.276)
2.014***
(0.337)
0.137***
(0.031)
29.3
0.006
0.002
0.000
5,817
0.823***
(0.264)
2.005***
(0.326)
0.126***
(0.028)
25.9
0.101
0.000
0.004
3.429***
(0.770)
1.490***
(0.184)
0.328***
(0.093)
16.5
0.012
0.027
0.022
1,939
2.619***
(0.310)
1.493***
(0.184)
0.323***
(0.081)
6.6
0.000
0.040
0.001
0.056
(0.603)
1.632***
(0.140)
0.224
(0.279)
17.5
0.193
0.710
0.006
1,939
0.717*
(0.392)
1.679***
(0.136)
−0.020
(0.188)
4.8
0.802
0.001
0.012
−0.319
(0.821)
1.324***
(0.390)
0.159**
(0.065)
65.2
0.077
0.092
0.000
31.344
(37.276)
2.121***
(0.596)
0.929
(0.918)
0.5
0.002
0.437
0.666
1,003
31.529
(37.075)
2.120***
(0.597)
0.931
(0.919)
0.4
0.001
0.432
0.728
1.202
(0.788)
1.929***
(0.255)
−0.169
(0.324)
16.9
0.533
0.417
0.438
1,003
1.092
(0.809)
1.928***
(0.255)
−0.139
(0.330)
26.3
0.595
0.361
0.360
Metropolitan areas, no Dallas, Houston, Oklahoma City, and Tulsa
DiffEnVar
0.653
0.745*
−0.963
(0.443)
(0.447)
(0.934)
DiffDShock
1.134***
1.134***
1.367***
(0.114)
(0.115)
(0.395)
Mining85
0.003
0.003
0.179***
(0.005)
(0.005)
(0.070)
Durbin P
161.6
1,568.0
23.0
Overid P
0.805
0.978
0.011
0.316
0.418
0.035
Wald Pa
IV F-stat
0.004
0.000
0.049
# Obs
11,033
3,009
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.)
Table 9
OLS estimation results for the subsamples of predominantly oil- and predominantly gas-producing nonmetropolitan counties.
Predominantly oil-producing
1-year diff
DV: total employment
DiffEnVar
1.181***
(0.112)
DiffDShock
1.198*
(0.595)
Mining85
0.005
(0.010)
0.308
R2
0.978
Wald Pa
# Clusters
12
# Obs
693
DV: employment in tradable industries
DiffEnVar
−0.022
(0.015)
DiffDShock
0.201
(0.196)
Mining85
−0.008**
(0.004)
0.0290
R2
0.279
Wald Pa
# Clusters
12
# Obs
693
Predominantly gas-producing
3-year diff
6-year diff
10-year diff
1-year diff
3-year diff
6-year diff
10-year diff
1.896***
(0.230)
2.684
(2.031)
0.050
(0.083)
0.586
0.695
2.310***
(0.363)
0.712
(1.444)
−0.086
(0.454)
0.810
0.255
1.939***
(0.249)
1.083
(0.649)
−0.495
(0.508)
0.800
0.285
1.841***
(0.474)
2.984**
(1.311)
0.015
(0.041)
0.501
0.372
1.734***
(0.339)
1.796***
(0.500)
−0.382*
(0.198)
0.507
0.900
1.029***
(0.166)
2.347***
(0.565)
0.274
(0.411)
0.512
0.012
189
63
63
1.155***
(0.188)
1.709***
(0.218)
−0.002
(0.004)
0.289
0.113
18
1,111
303
101
101
0.131**
(0.054)
1.633
(1.553)
−0.049**
(0.019)
0.0659
0.344
−0.079
(0.096)
0.338
(0.609)
−0.010
(0.188)
0.111
0.503
−0.011
(0.056)
0.870**
(0.335)
0.172
(0.250)
0.687
0.0296
−0.059
(0.072)
0.860*
(0.421)
0.000
(0.020)
0.187
0.0477
−0.140
(0.093)
0.817**
(0.298)
−0.115
(0.087)
0.269
0.00462
−0.122
(0.070)
0.010
(0.229)
0.112
(0.128)
0.697
0.622
189
63
63
0.059
(0.068)
0.940***
(0.106)
−0.003*
(0.002)
0.172
1.28e-06
18
1,111
303
101
101
1.389***
(0.293)
0.374
(1.144)
−0.076
(0.359)
0.678
0.358
1.095***
(0.207)
0.101
(0.554)
−0.200
(0.333)
0.653
0.119
0.901*
(0.496)
2.124
(1.279)
0.015
(0.042)
0.381
0.308
0.875**
(0.387)
0.979
(0.610)
−0.266
(0.190)
0.264
0.856
0.444**
(0.187)
1.835***
(0.445)
0.448
(0.325)
0.414
0.00212
63
63
0.096
(0.212)
0.769***
(0.183)
0.001
(0.003)
0.122
0.0319
18
1,111
303
101
101
DV: employment in non-tradable industries
DiffEnVar
0.203*
0.765***
(0.104)
(0.190)
DiffDShock
0.997
1.051
(0.615)
(1.622)
Mining85
0.013
0.098
(0.008)
(0.070)
0.0724
0.327
R2
0.209
0.860
Wald Pa
# Clusters
12
# Obs
693
189
***, **, * significant at 0.01, 0.05, and 0.1, respectively; standard errors adjusted for clustering at BEA region level in parentheses; all models include a full set of controls and time period
dummies described in text and note to Table 1. aP-value from the Wald test of the coefficients equality (H0: βDiffEnVar = βDiffDShock.).
94
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
OLS results for total, tradable, and nontradable sectors separately by
subsamples of nonmetro counties.
Overall, the results for the spillovers from energy extraction are consistent with the base ones in Tables 1, 3, and 4, though the six-year multiplier effects are smaller in this case. The positive effects of the energy
variable DiffEnVar in the total and non-tradable employment models
are comparable to the entire nonmetro sample estimates and follow approximately the same pattern over time. For the employment in tradable industries, the results are mostly insignificant. As in our main
results, the average industry multiplier in the natural gas sample
tends to exceed the energy multiplier, but this is not the case in the predominantly oil producing counties. Given the very small sample, the results should be taken with caution, although a story of energy shocks
having larger effects in remote heavy oil producing and very sparsely
populated regions in the states of New Mexico, North Dakota, West
Texas, and Wyoming seems plausible. One still needs to keep in mind,
however, that the results of the Wald test (perhaps somewhat unreliable due to the small number of observations) in most cases point to
the equality of the coefficients on the two main explanatory variables,
though the SUR regression coefficients suggested otherwise.
points to the need to exercise caution when designing policies. Yet,
we believe that the local employment spillovers from oil and gas development tend to be more modest than equal-sized shocks from other industries (on average), suggesting less complementarities in the
economy. Population responses to both types of shocks are small, suggesting limitations in the role of agglomeration economies supporting
long-term growth in modern booms. Given that energy booms are
small in magnitude relative to the rest of the economy, local economies
would appear to be better off when they experience broad-based
growth rather than the energy-led booms both in terms of multiplier effects and in terms of enhancing the diversity of their economies.
Policymakers who look at oil and gas development as a silver bullet
for their economic development woes should probably consider other
sectors, which on average appear to have larger local employment
effects.
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://dx.
doi.org/10.1016/j.eneco.2016.07.015.
6. Conclusions
As observed by others (Munasib and Rickman, 2015; Weber, 2012),
the impact of the energy sector on various economic performance metrics is uneven across space, time, and industrial landscape. Our analysis
finds considerable heterogeneity in energy sector effects on employment in metropolitan and nonmetropolitan US counties between
1993 and 2013. The responses also vary across the different time horizons we consider. Even instrument choice can sometimes affect the
coefficients.
Estimation results suggest that in nonmetropolitan areas, the
energy-sector multiplier on total county employment first increases
and peaks at six years decreasing afterwards. Only at six years is the typical multiplier response to oil and gas shocks larger than the corresponding response to the average economy-wide shock. Positive
energy-sector spillovers to other industries are most pronounced in
construction, transportation and warehousing, wholesale trade, accommodation, real estate, and nontradable goods sector. There appears to
be some crowding out of traded goods sectors by oil and gas booms,
consistent with the Dutch Disease hypothesis. Indeed, even in industries
that are close upstream suppliers of the energy sector, oil and gas expansion does not appear to statistically stimulate employment gains
until about ten years later, and even then, it is no larger than what an
equally-sized shock in other industries would do. Hence, small populations in nonmetropolitan areas do not have the size or scale to support
oil and gas supply-chain activities, or alternatively, because these industries have long established themselves in the regions of the oil patch,
there is less need to create new firms elsewhere (Brown and Lambert,
2016).
For metropolitan areas, there is no statistical relationship between
oil and gas industry expansion and total employment. At disaggregated
level, only a few sectors respond to changes in oil and gas extraction in
the metro sample with weakly positive effects observed for some
differencing spans in the equations of employment in non-tradable industries, construction, agriculture, transportation and warehousing.
Wholesale trade is the only sector where the stimulating impact of energy sector is significant at the 0.05 level in the 3-year and 10-year
differenced models.
Overall, our results point to the importance of a number of factors,
such as the sample, choice of instruments, industries of interest and
timeframe that should be explicitly acknowledged when discussing energy sector impacts. Arguably, the accumulated evidence to a considerable extent depends on these factors. This makes generalizations of the
existing findings in the literature problematic and may explain why extant research comes to conflicting conclusions. For policymakers, this
References
Alexeev, M., Conrad, R., 2009. The elusive curse of oil. Rev. Econ. Stat. 91 (3), 586–598.
Allcott, H., Keniston, D., 2014. Dutch Disease or agglomeration? The local economic effects
of natural resource booms in modern America. Retrieved from http://www.nber.org/
papers/w20508.
Angrist, J.D., Pischke, J.-S., 2008. Mostly Harmless Econometrics: An Empiricist's Companion. Princeton University Press.
Angrist, J.D., Pischke, J.-S., 2014. Mastering' Metrics: The Path from Cause to Effect.
Princeton University Press.
Bartik, T.J., 1991. Who Benefits from State and Local Economic Development Policies?
Books from Upjohn Press.
Betz, M.R., Farren, M., Lobao, L., Partridge, M., 2015. Coal mining, economic development,
and the natural resource curse. Energy Econ. 50, 105–116.
Bhattacharyya, S., Hodler, R., 2010. Natural resources, democracy and corruption. Eur.
Econ. Rev. 54 (4), 608–621.
Bjorvatn, K., Farzanegan, M.R., Schneider, F., 2012. Resource curse and power balance:
evidence from oil-rich countries. World Dev. 40 (7), 308–1316.
Black, D., McKinnish, T.G., Sanders, S.G., 2005. Tight labor markets and the demand for education: evidence from the coal boom and bust. Ind. Labor Relat. Rev. 59 (1), 3–16.
Blanco, L., Grier, R., 2012. Natural resource dependence and the accumulation of physical
and human capital in Latin America. Res. Policy 37 (281-295).
Boxall, P., Chan, W., McMillan, M., 2005. The impact of oil and natural gas facilities on rural
residential property values: a spatial hedonic analysis. Resour. Energy Econ. 27,
248–269.
Boyce, J.R., Emery, J.C. Herbert, 2011. Is a negative correlation between resource abundance and growth sufficient evidence that there is a “resource curse”? Res. Policy
36 (1), 1–13.
Brown, J., 2014. Production of natural gas from shale in local economies: a resource blessing or curse? Econ. Rev. Q I, 1–29.
Brown, J., Lambert, D.M., 2016. Extending a smooth parameter model to firm location
analyses: the case of natural gas establishments in the United States. J. Reg. Sci.
http://dx.doi.org/10.1111/jors.12280 (n/a-n/a).
Brunnschweiler, C.N., Bulte, E.H., 2008. The resource curse revisited and revised: a tale of
paradoxes and red herrings. J. Environ. Econ. Manag. 55 (3), 248–264.
Carmignani, F., 2014. The Spatial Curse of Natural Resources (Retrieved from).
Cavalcanti, T.V.d.V., Mohaddes, K., Raissi, M., 2011. Growth, development and natural resources: new evidence using a heterogeneous panel analysis. Q. Rev. Econ. Finance
51, 305–318.
Dorfman, J.H., Partridge, M.D., Galloway, H., 2011. Do natural amenities attract high-tech
jobs? Evidence from a smoothed Bayesian spatial model. Spat. Econ. Anal. 6 (4),
397–422. http://dx.doi.org/10.1080/17421772.2011.610811.
Duranton, G., Turner, M.A., 2011. The fundamental law of road congestion: evidence from
US cities. Am. Econ. Rev. 101 (6), 2616–2652.
Fallah, B., Partridge, M., Olfert, M.R., 2011. New economic geography and US metropolitan
wage inequality. J. Econ. Geogr. 11 (5), 865–895.
Fallah, B., Partridge, M., Rickman, D.S., 2014. Geography and high-tech employment
growth in US counties. J. Econ. Geogr. 14 (4), 683–720.
Fleming, D.A., Measham, T.G., 2014. Local job multipliers of mining. Res. Policy 41 (C),
9–15.
Frankel, J.A., 2010. The Natural Resource Curse: A Survey (Retrieved from).
Freeman, D.G., 2009. The ‘resource curse’ and regional US development. Appl. Econ. Lett.
16 (5), 527–530.
Glaeser, E.L., Gottlieb, J.D., 2008. The Economics of Place-Making Policies (Retrieved
from).
Gylfason, T., 2000. Resources, agriculture, and economic growth in economies in transition. Kyklos 53 (4), 337–361.
A. Tsvetkova, M.D. Partridge / Energy Economics 59 (2016) 81–95
Gylfason, T., 2001. Natural resources, education, and economic eevelopment. Eur. Econ.
Rev. 45 (4-6), 847–849.
Gylfason, T., Zoega, G., 2014. The Dutch Disease in reverse: Iceland's natural experiment.
Retrieved from http://EconPapers.repec.org/RePEc:oxf:oxcrwp:138.
Gylfason, T., Herbertsson, T.T., Zoega, G., 1999. A mixed blessing: natural resources and
economic growth. Macroecon. Dyn. 3 (02), 204–225.
Haggerty, J., Gude, P.H., Delorey, M., Rasker, R., 2014. Long-term effects of income specialization in oil and gas extraction: the U.S. west, 1980–2011. Energy Econ. 45 (0),
186–195.
Jacobsen, G., Parker, D., 2014. The economic aftermath of resource booms: evidence from
boomtowns in the American West. Econ. J.
James, A., Aadland, D., 2011. The curse of natural resources: an empirical investigation of
U.S. counties. Resour. Energy Econ. 33 (2), 440–453.
James, A., James, R.G., 2011. Do resource dependent regions grow slower than they
should? Econ. Lett. 111 (3), 194–196.
Kelsey, T.W., Partridge, M.D., White, N.E., 2016. Unconventional gas and oil development
in the United States: economic experience and policy issues. Appl. Econ. Perspect.
Policy 38 (2), 191–214. http://dx.doi.org/10.1093/aepp/ppw005.
Kinnaman, T., 2011. The economic impact of shale gas extraction: a review of existing
studies. Ecol. Econ. 70 (7), 1243–1249.
Lewbel, A., 2012. Using heteroscedasticity to identify and estimate mismeasured and endogenous regressor models. J. Bus. Econ. Stat. 30 (1), 67–80.
Manzano, O., Rigobon, R., 2001. Resource curse or debt overhang? National Bureau of
Economic Research Working Paper Series, No. 8390.
Marchand, J., 2012. Local labor market impacts of energy boom–bust–boom in Western
Canada. J. Urban Econ. 71 (1), 165–174.
Mehlum, H., Moene, K., Torvik, R., 2006. Institutions and the resource curse. Econ. J. 116
(508), 1–20.
Michaels, G., 2011. The long term consequences of resource-based specialisation. Econ. J.
121 (551), 31–57.
Munasib, A., Rickman, D., 2015. Regional economic impacts of the shale gas and tight oil
boom: a synthetic control analysis. Reg. Sci. Urban Econ. 50 (1), 1–17.
Murshed, S.M., Serino, L.A., 2011. The pattern of specialization and economic growth: the
resource curse hypothesis revisited. Struct. Chang. Econ. Dyn. 22 (2), 151–161.
Nolan, C., Morrison, E., Kumar, I., Galloway, H., Cordes, S., 2011. Linking industry and occupation clusters in regional economic development. Econ. Dev. Q. 25 (1), 26–35.
Papyrakis, E., Gerlagh, R., 2007. Resource abundance and economic growth in the United
States. Eur. Econ. Rev. 51 (4), 1011–1039.
95
Paredes, D., Komarek, T., Loveridge, S., 2015. Income and employment effects of shale gas extraction windfalls: evidence from the Marcellus region. Energy Econ. 47 (0), 112–120.
Partridge, M., Rickman, D., 2006. An SVAR model of fluctuations in U.S. migration flows
and state labor market dynamics. South. Econ. J. 72 (4).
Partridge, M., Rickman, D.S., Olfert, M.R., Ali, K., 2012. Dwindling US internal migration:
evidence of spatial equilibrium or structural shifts in local labor markets? Reg. Sci.
Urban Econ. 42 (1), 375–388.
Peach, J., Starbuck, M., 2011. Oil and gas production and economic growth in New Mexico.
J. Econ. Issues XLV (2), 511–526.
Ross, M.L., 1999. The political economy of the resource curse. World Polit. 51 (02), 297–322.
Sachs, J., Warner, A., 1999. The big push, natural resource booms and growth. J. Dev. Econ.
59 (1), 43–76.
Sachs, J., Warner, A., 2001. The curse of natural resources. Eur. Econ. Rev. 45,
827–838.
Sala-i-Martin, X., Subramanian, A., 2013. Addressing the natural resource curse: an illustration from Nigeria. Journal of African Economies 22(4). Centre for the Study of
African Economies (CSAE), pp. 570–615.
Sala-I-Martin, X., Doppelhofer, G., Miller, R.I., 2004. Determinants of long-term growth: a
Bayesian averaging of classical estimates (BACE) approach. Am. Econ. Rev. 94 (4),
813–835.
Sargan, J.D., 1958. The estimation of economic relationships using instrumental variables.
Econometrica 393–415.
Sovacool, B., 2014. Cornucopia or curse? Reviewing the costs and benefits of shale gas
hydraulic fracturing (fracking). Renew. Sust. Energ. Rev. 37, 249–264.
Stock, J.H., Wright, J.H., Yogo, M., 2002. A survey of weak instruments and weak identification in generalized method of moments. J. Bus. Econ. Stat. 20 (4).
Van der Ploeg, F., 2011. Natural resources: curse or blessing? J. Econ. Lit. 366–420.
Weber, J.G., 2012. The effects of a natural gas boom on employment and income in
Colorado, Texas, and Wyoming. Energy Econ. 34 (5), 1580–1588.
Weber, J.G., 2014. A decade of natural gas development: the makings of a resource curse?
Resour. Energy Econ. 37 (0), 168–183.
Weinstein, A., 2014. Local labor market restructuring in the shale boom. J. Reg. Anal. Policy
44 (1), 71–92.
Wooldridge, J.M., 2010. Econometric Analysis of Cross section and Panel Data. MIT Press.
Wooldridge, J.M., 2013. Introductory Econometrics: A Modern Approach. fifth ed. Cengage
Learning, New Delhi.