Skip to main content

Abstract

Oil and gas windfalls may lead to the Dutch disease, that is, the crowding out of the manufacturing sector due to rising wages when labor is drawn to the expanding sectors. In this paper, we exploit the fact that oil and gas discoveries contain an element of luck as well as oil price fluctuations to capture exogenous variation in oil and gas windfalls across Indonesia and identify their effects on manufacturing firms. We find that oil and gas windfalls on average cause wages as well as firms’ labor productivity, output, and employment to increase, while product unit values and exit rates are unaffected. Heterogeneity analysis reveals that the least productive firms are more likely to exit, and surviving low-productivity firms see relatively large expansions in output and labor productivity, while high-productivity firms see relatively high expansions in employment.

Whether natural resource wealth is a curse or a blessing is still subject to intense debate among economists. Over the last two decades, many resource-rich developing countries have seen strong economic growth, but the crash in the oil price in 2014 exposed the fragility of economies highly dependent on resource-based exports and government revenues. While price booms and new discoveries can drive overall economic expansion, there remain concerns that the manufacturing sector can be crowded out, otherwise known as Dutch disease (Corden and Neary 1982).

In this paper, we study the local effects of oil and gas booms on manufacturing firms in Indonesia. Our identification strategy is to compare firms across districts (kabupatens) that differ in terms of yearly oil and gas windfalls. Detailed data on oil and gas exploration allow us to calculate long-term district-specific oil and gas discovery rates, that is, the share of nondry wells among exploration wells. The idea is that discoveries, conditional on exploration drilling, are determined by geology and luck and therefore provide exogenous variation in oil and gas exposure across districts. By combining the discovery rates with the world oil price, we generate plausibly exogenous variation in oil and gas windfalls across districts and years.

Within-country analysis helps in separating the effects of natural resources from other macroeconomic variables. For example, openness to international trade, quality of institutions, and financial development might confound the relationship between resources and manufacturing at the cross-country level (van der Ploeg 2011). The use of firm-level data allows us to focus on the microfoundations of the Dutch disease theory. Dutch disease within a country may be triggered by booms increasing local factor demand (Allcott and Keniston 2017). In our setting, a boom may occur in the resource sector, the services sector, or the public sector. It is well documented that higher oil prices increase the activity in the oil sector (Anderson et al. 2018), while windfalls have been found to expand the nontraded goods sector (Harding et al. 2016; Benguria et al. 2018). With increased labor demand from the resource sector or the nontraded sectors, local manufacturing firms may face increased wages due to upward-sloping local labor supply and migration frictions.1 In contrast to the price of services, the price of manufactured goods is anchored by within-country trade and international trade.2

Indonesia provides an ideal setting for our study, as it is a relatively large developing country with a sizable manufacturing sector and many regions producing oil and gas. Resource rents have consistently accounted for more than 10% of GDP and 20% of exports over the last 20 years. Manufacturing exports have typically counted for more than one-third of total exports, consisting mostly of textiles, plastics, and rubbers, as well as machinery and electronics. To the best of our knowledge, this is the first paper providing evidence on the effects of oil and gas windfalls on manufacturing firms in an emerging economy.

We find that a 10% increase in windfalls causes an overall 1.8% increase in wages and no increase in product prices. The wage effect appears predominantly via the wages of production workers. Further we find that windfalls raise output, employment, and labor productivity. Regarding firm entry, the introduction of new products, and dropping of existing products, we find little effects of windfalls. Allowing the effect of oil and gas windfalls to depend on four dimensions of firm heterogeneity, our results reveal stark differences across firms.

First, we find that firms with low initial labor productivity either grow or die in the wake of a windfall. They are significantly more likely to exit, while those surviving show larger expansions in output, labor productivity, and the use of intermediate inputs than their initially more productive peers. They also raise the wages they pay and the unit values of their products more than other firms do.

Second, zero initial export share plays about the same role as low initial labor productivity, as nonexporting firms expand output, labor productivity, and use of intermediate inputs more than their exporting peers. Export-oriented firms are more likely to drop some products, and they reorient their production away from exports in the sense that the share of their production that is exported decreases. Also export-oriented firms expand their output, employment, and labor productivity, and only the few firms with very high export shares show significant output contractions. Consistent with prices being anchored in the international market, export orientation holds back wage increases and unit value increases.

Third, low initial skill intensity contributes to decreasing employment and investments in machinery, the latter being indicative of complementarity between skill intensity and capital. Low skill intensity also works to reduce the wage increase of production workers, that is, the workers these firms hire relatively intensively. For the nonproduction workers, on the other hand, low skill intensity of the firms works to increase wages compared to nonproduction workers in more skill intensive firms.

Fourth, firms with high initial labor intensity expand their output and their use of intermediate inputs more than other firms. They experience a larger boost in labor productivity, as they expand employment less than other firms. They raise the wages of their production workers less, just as the firms with low skill intensity and consistent with limited room to expand the unit price of a factor they use intensively. Finally, they see increases in the unit values of their products compared to other firms, which could indicate a more local orientation of these firms. The positive effect on unit values and the larger quantity produced are consistent with a positive demand shock facing firms with relatively high initial labor intensity.

Our results are similar to Michaels (2011), Cavalcanti et al. (2019), Allcott and Keniston (2017), Feyrer et al. (2017), and Benguria et al. (2018) in that wages and domestic demand are found to increase in response to a commodity boom. This would also be in line with the spending effect from the standard theory on the Dutch disease: windfalls increase wages as local demand, notably in the nontradable sectors, puts pressure on labor markets (Corden and Neary 1982). However, the standard theory prediction of a labor reallocation away from manufacturing (due to increased wages and held-down prices) is not supported in our data. Instead, we find that manufacturing firms on average manage to grow in terms of output and employment, even without raising the unit values of their products.3

Our findings suggest that windfalls lead to aggregate productivity improvements: the least productive firms exit, surviving low productivity firms improve their labor productivity, and the most productive firms expand their employment. These productivity results are in line with Stefanski and Kuralbayeva (2013), who find in cross-country macrodata, as well as in US county microdata, that manufacturing sectors in resource-rich regions are small but have relatively high productivity. Consistent with our finding on exit, their model and Melitz (2003) highlight selection effects, where the least productive exit from manufacturing when faced with a cost shock. Our finding that surviving firms on the left of the productivity distribution catch up and grow more than their larger and more productive competitors is also in line with the prediction of Melitz (2003). Our estimated positive effects on labor productivity, as well as the finding of Allcott and Keniston (2017) of no significant differences in total factor productivity in any sector due to windfalls, suggest that the spending effect does not necessarily lead to deteriorating productivity.4

We do not find evidence of contraction among export-oriented firms due to windfalls, except for the few firms with very high export shares. Allcott and Keniston (2017)find a contraction in the manufacturing sector producing tradable goods during resource booms for the United States. Harding and Venables (2016) find evidence of contracting nonresource exports in response to resource-trade windfalls but show that the contracting effect on nonresource trade is more pronounced for countries with good institutions, high GDP per capita, and goods with lower trade costs (a proxy for how footloose the sector may be). Our seemingly deviating results with respect to export orientation may be explained by the fact that we study a large developing country where much of the manufacturing industry may serve the domestic market.

This paper contributes to the literature on the resource curse by showing that local resource shocks are not detrimental for manufacturing in a large developing country. Allcott and Keniston (2017) show that manufacturing firms located in US counties with large oil and gas endowments did not suffer from Dutch disease effects. We find the same result in Indonesia, in spite of the expectation that manufacturing may be more labor intensive and more oriented toward low-cost production in developing countries.5 We move beyond Allcott and Keniston (2017) in that we provide evidence on unit values as well as on wages for production workers and nonproduction workers separately. We also investigate in detail the margins along which the firms adjust, including exit and entry, product drops and introduction of new products, the share of production exported, the use of intermediate inputs, the share of the intermediate inputs that are imported, and investments in machines. We move beyond heterogeneity across sectors, which has been the standard approach also in the few papers using firm-level data (e.g., Allcott and Keniston 2017; De Haas and Poelhekke 2019) and investigate firm heterogeneity within sectors with respect to productivity, export orientation, labor intensity, and skill intensity.6

Our identification strategy differs from the rest of the literature as it is based on within-firm analysis and utilizes district-specific luck and geology in combination with the world price of oil as plausible exogenous variation in resource windfalls. This strategy has several advantages. First, it follows Dube and Vargas (2013), Bazzi and Blattman (2014), and Harding and Venables (2016) in using world market prices for time variation. This helps us with mimicking variations in windfalls due to price changes and not production changes. What typically happens is that production starts with a lag after a discovery is made and follows some extraction path which mainly depends on geological aspects of the reservoirs; see, for example, Anderson et al. (2018) and Arezki, Ramey, and Sheng (2017). Price booms, on the other hand, are quite different, creating large fluctuations in resource rents and following economic adjustment challenges. Second, it deals with the main empirical challenge with using prices as the windfall measure, which is that the weights across districts are potentially endogenous. The discovery rate across districts is determined by geology and luck rather than other circumstances potentially correlated with the performance of the manufacturing sector.7 Our strategy has similarities with the use of Bartik-like instruments, and it is central for identification that the weights determining each district’s exposure to the time variation are exogenous. We show that the district-specific discovery rates do not correlate with potential confounders, as suggested by Goldsmith-Pinkham et al. (2018).8 Third, the detailed manufacturing data allow us to control for unobservable time-invariant firm characteristics, such as firm location, with firm fixed effects and sector-year-specific shocks, such as demand shocks, with sector-year fixed effects.

The paper proceeds as follows. In section 1, we describe the empirical strategy, our data, and the setting of oil and gas production in Indonesia. In section 2, we present our results, and in section 3 we conclude.

1.  Data and Empirical Strategy

1.1.  Empirical Strategy

Oil and gas extraction may take place where institutions and infrastructure are of higher relative quality and, thus, where manufacturing is also favored. To identify the effect of windfalls on manufacturing firms, we need exogenous variation in the location and timing of windfalls. Our strategy is to proxy oil and gas windfalls with the interaction of luck in oil and gas discovery in the district with yearly variation in oil prices. The number of wells drilled in a district may be determined by district-specific factors such as infrastructure and institutions, consistent with the finding of Cust and Harding (2019) that the amount of exploration across countries is an outcome of institutional quality. The discovery rate, that is, the share of nondry wells among exploration wells, is in contrast not found to vary with institutional quality across countries. As we show later in subsection 1.3, the same is true across districts in Indonesia. We use the mean district-specific discovery rate in our period in the regressions.

The discovery rate is likely to be determined by geology and luck.9 Yearly variation in the oil price, which is set on the world market, can thus be combined with variation in luck across districts to generate exogenous variation in the location and timing of windfalls.10

We estimate the effect of oil and gas windfalls on firm-level outcomes across districts and years by the following regression:

(1)Yit=αi+δjt+βLi×Pt+uit,
where Li is luck, that is, the number of nondry wells over total wells drilled in the district of firm i, Pt is the oil price in year t in 2010 USD, αi and δjt are firm- and industry-year fixed effects. The former helps in controlling for firm-specific time-invariant unobservables, such as location and product brand, whereas the latter helps in controlling for time-varying shocks hitting specific sectors, such as global demand shocks. The unit of observation is the firm-year. We cluster standard errors at the district level to take into account potential serial correlation and to adjust for potential Moulton bias, as the treatment is observed at a more aggregate level than the outcome.

Our treatment intensity is captured by variation in luck in discovery across districts and in yearly variation in the oil price. Therefore, we must drop all districts that never explored for oil and gas. This is similar to the identification strategy of Cavalcanti et al. (2019), who suggest that Brazilian districts with exploration but no discovery are a valid counterfactual to estimate the effect of discoveries. We improve on this approach by using the ratio of exploration success to drilling per district instead of a dummy and, importantly, by using the oil price as a source of time variation rather than simply a dummy that remains one for all years postdiscovery.

We run parsimonious models with no control variables. The reason for this choice is that, with proper random variation, the role of the controls would be just to increase the precision of the estimates. However, including controls that are themselves potentially affected by the treatment makes the interpretation more difficult.

1.2.  Data

Well-level exploration and production data are provided by the PathFinder database owned by Wood Mackenzie (2011). In Indonesia, it includes more than 1,200 individual wells in over 40 districts (out of a total of 280 districts, known as kabupatens).11 Data on the oil price are from the BP Statistical Review of World Energy 2016.

The data on firms are from the Indonesian Manufacturing Census for the years 1990–2008. They cover all plants with more than 20 employees, that is, more than 30,000 plants during 1990–2008. These data provide us with information about wages, exit rates, employment, productivity, and output prices that we can compute using core products’ unit values for which we have data from 1998 onward. The distribution of sector of activity is given in table 1.

Table 1. 

Two-Digit Industries, ISIC Revision 3

ISIC Two-DigitISIC Two-Digit NameN
15Manu. of food products and beverages7,617
16Manu. of tobacco products1,544
17Manu. of textiles3,646
18Manu. of wearing apparel; dressing and dyeing of fur3,112
19Tanning and dressing of leather; manu. of luggage, handbags, saddlery, harness, and footwear922
20Manu. of wood and of products of wood and cork, except furniture; articles of straw and plaiting material2,186
21Manu. of paper and paper products709
22Publishing, printing, and reproduction of recorded media990
23Manu. of coke, refined petroleum products, and nuclear fuel232
24Manu. of chemicals and chemical products1,384
25Manu. of rubber and plastics products1,869
26Manu. of other nonmetallic mineral products2,579
27Manu. of basic metals406
28Manu. of fabricated metal products, except machinery and equipment1,218
29Manu. of machinery and equipment n.e.c.650
30Manu. of office, accounting, and computing machinery266
31Manu. of electrical machinery and apparatus n.e.c.427
32Manu. of radio, television, and communication equipment and apparatus380
33Manu. of medical, precision, and optical instruments, watches, and clocks240
34Manu. of motor vehicles, trailers, and semi-trailers425
35Manu. of other transport equipment453
36Manu. of furniture; manufacturing n.e.c.3,122
37Recycling86
 Total 34,463

Note. ISIC = International Standard Industrial Classification. “n.e.c.” = not elsewhere classified.

View Table Image

We follow Amiti and Davis (2012) and use the log of the average firm-level wage, defined as the total wage bill divided by the number of workers. We define the exit rate for each period as the share of firms we no longer observe in the subsequent period. Employment is simply the number of employees, which we can categorize as white or blue collar, labeled as production/nonproduction workers in the data. We choose a simple ratio of output per employee as our measure of labor productivity. To proxy for output prices we use the unit values (value/volume) of the firm’s core product, defined as the one with maximum value.

1.3.  The Setting: Indonesia

Indonesia first discovered oil in 1885 in North Sumatra. PWC (2014) suggests that in 2014 it accounted for 1.1% of world oil production. The country’s main oil-producing areas include Sumatra, the Java Sea, and East Kalimantan. The same report estimates that Indonesia produced 2.1% of the world natural gas in 2012, making it the fourth largest natural gas producer. The main gas-producing regions include South Sumatra and East Kalimantan. McKinsey & Co. (2014) says Indonesia has 22 billion barrels of conventional oil and gas reserves, of which 4 billion are recoverable. That is the equivalent of about 10 years of oil production and 50 years of gas. Nonetheless, given domestic consumption levels, Indonesia has been a net importer of oil since 2004.

Indonesia is considered a resource-rich economy (IMF 2012), and the left-hand panel in figure 1 shows its high share of resource rents in GDP. It has experienced erratic oil production despite a gas boom and an oil price boom in the 2000s (fig. 1, right-hand panel). Figure 2 shows a map of oil and gas production by district in Indonesia. Oil production and exploration is spread out across the country; manufacturing firms at different locations are hence exposed to different intensities of oil and gas windfalls.

Figure 1. 
Figure 1. 

Resource rents and production of oil and gas: resource rents in the left panel, and production of oil and gas in Indonesia in the right panel. Source: Resource rents from World Development Indicators, World Bank, and oil and gas production from the PathFinder database, Wood Mackenzie (2011).

Figure 2. 
Figure 2. 

Oil and gas production across districts in Indonesia. Around 85% of districts have no oil or gas production in any period; 35% of production fields are offshore but can be linked to a closest district.

In figure 3 we describe our exploration-luck variable to show that it can indeed capture oil and gas windfalls, being correlated with oil and gas production and government revenues, while being orthogonal to infrastructure and institutions.12 The top panel shows the distribution of success rates across district years. For around half of our districts the success rate is 100%. Yet around 12% of district-years have zero luck, with a success rate of 0. Success in other district-years varies between 20% and 90%. The top right figure shows how average luck is relatively stable across years, even in times of high prices where we might expect discoveries to have higher payoffs and companies wanting to discover more. The graph indeed shows that luck does not appear to respond to or vary systematically with prices. In the second row, the right figure shows that drilling a large number of exploration wells in a particular district in a particular year does not translate into higher, or lower, success rates. Across district-years, the slope is close to zero. If exploration indeed captures the endogenous aspect of oil windfalls, it is reassuring to confirm that luck is not a function of exploration wells. The left figure in the second row shows that average luck on the other hand is positively correlated with oil and gas production across districts during our study period. In the third row, we show that luck is positively correlated with government revenues, as is oil and gas production, given Indonesia’s extensive degree of fiscal decentralization. This confirms that we are able to capture variation in windfalls across districts by using luck in exploration. This is important, as in our context the spending effect is likely linked to windfalls in public funds in oil districts and thus to greater local demand for goods and services. Finally, the two bottom figures confirm that average luck is orthogonal to infrastructure quality, measured here by travel time to major cities, and to institutional quality, measured here with an index of economic governance.

Figure 3. 
Figure 3. 

Luck in exploration. Success rate is defined as between 1990 and 2008, unless when it is specified for 1998–2008. It is defined as nondry wells over total exploration wells. Source: Oil and gas exploration wells from the PathFinder database, Wood Mackenzie (2011). Government revenues are from the Ministry of Finance Regional Financial Information System. Travel time data are from Nelson and Uchida (2008). Economic and Governance Index is from the Asia Foundation. Note that in the right scatter in the upper middle panel we have observations at the district-year level and hence cluster standard errors by district, as in our regression. The other scatters are averaged cross-sections.

2.  Effect of Oil and Gas Windfalls

2.1.  Within-Country Trends

To understand the trends within Indonesia we provide two sets of comparisons. First, we compare macrotrends between districts with high exposure to oil and districts with low exposure to oil. Our preferred comparison is between lucky and unlucky districts, that is, we exclude all districts that have not drilled for oil. Comparing unlucky with lucky mimics our identification strategy and provides a more convincing counterfactual than comparing municipalities in terms of their oil and gas production. For these figures, we consider districts lucky if they drilled and found oil, that is, if their discovery rate is above 0% (in the regressions we instead use discovery rates as a continuous variable). For completeness, we compare also districts with oil and gas production versus districts without such production. Figure 4 presents the trends for wages, exit rates, and output.

Figure 4. 
Figure 4. 

Wages (A), exit (B), and output (C). Average wage is the mean across firms of our wage variable at the firm level, which is defined as (total wage sum)/(total employment). Exit rate is the share of firms that exit that year, that is, for which this is the last year we observe them in the data. Manufacturing output is the sum of all firms’ output. Luck in exploration means an average discovery rate above 0%. No Luck means an average discovery rate of 0%. If there is no exploration, there is no discovery rate. Oil and Gas are districts that have oil or gas production at some point. No Oil and Gas have zero production over the whole period. Note that No Oil and Gas together with Oil and Gas include all districts in Indonesia. The “naive” way is to compare oil and gas districts to all the other ones (left-hand-side panels), and our preferred way is to compare the lucky ones with the unlucky ones among districts with drilling (right-hand-side panels).

Panel A of figure 4 suggests that oil and gas districts have been associated with higher wages on average across Indonesia, as predicted by the standard Dutch disease theory. The rate of wage growth, as seen from the slopes, also appears slightly higher when we compare firms in lucky and unlucky districts.

In panel B of figure 4 we compare the average exit rates across the same comparison groups. We find no systematic differences in exit rates between districts with and without oil production. Between lucky and unlucky districts we find that exit rates trend very differently at the beginning of the sample but follow a similar pattern from 1995 onward. This is also what appears when we compare the number of firms across the type of districts (not shown to save space). None of the graphs on exit rates suggest that oil and gas production has forced manufacturing firms out of business, despite the Dutch disease symptom of higher wages.13

In panel C of figure 4 we show that there is a distinguishable increase in total output in oil- and gas-producing and lucky districts during the period 1997–2000. This coincides with the growth in wages observed in figure 4 as well as with a period of growth in the oil price. It thus suggests that manufacturing firms may have expanded when oil and gas windfalls increased.

The data suggest that wage growth has been on average 3% higher in oil and gas districts. As a first test for the statistical significance of this relationship, we regress average wage growth on average oil and gas production across districts, allowing for heterogeneity across industries. Figure 5 presents the 16 sectors out of 23 for which we estimate a positive effect of oil and gas production on wages. While the confidence intervals are large, the figure indicates that in some sectors, firms in lucky districts saw faster wage growth.

Figure 5. 
Figure 5. 

Wages grew faster in lucky districts. Cross-section of firms. Wage growth is the average yearly wage growth over 1990–2006. Exploration luck is the mean success rate of drilling per district.

These descriptive results are consistent with the theoretical literature on the Dutch disease in that they suggest that oil and gas windfalls are associated with higher wage growth, yet we do not observe a crowding out of the manufacturing sector. In the next section, we explore the within-firm effect on wages and other outcomes with panel regressions.

2.2.  Firm-Level Evidence

2.2.1.  Wages and Unit Values

To estimate the effect of oil and gas windfalls on firms, we follow the model and identification strategy described in section 1. Results are presented in table 2, which includes reduced-form panel estimates focusing only on districts where there has been some oil and gas exploration. Note that we always include firm fixed effects and industry-year fixed effects; that is, the estimates are based on within-firm and within-industry-year variation. In other words, they do not reflect reallocations across firms or changes to industries over time.

Table 2. 

Wages and Unit Values

 Wage
(1)
Wage Production
(2)
Wage Nonproduction
(3)
Unit Value
(4)
 A1. Mean Effect across All Firms
Average success rate × oil price.175***.245***.081.070
 (.056)(.059)(.274)(.098)
N49,79549,79549,79548,530
R-squared.79.90.80.84
 A2. Mean Effect across All Firms, Controlling for Time Trend Interacted with Success Rate
Average success rate × oil price.175***.245***.082.070
 (.057)(.060)(.274)(.098)
Average success rate × time trend−.000−.000.000−.000
 (.000)(.000)(.000)(.000)
N49,79549,79549,79548,530
R-squared.79.90.80.84
 A3. Mean Effect across Firms Existing in 1990
Average success rate × oil price.186***.243***−.041.030
 (.053)(.061)(.271)(.107)
N38,71338,71338,71337,897
R-squared.79.91.80.84
 B. Heterogeneous Effects, Depending on Initial Labor Productivity
Average success rate × oil price1.446***1.059***.8691.592***
 (.197)(.196)(.546)(.392)
× 1990s labor productivity−.139***−.090***−.101*−.173***
 (.021)(.022)(.059)(.039)
N38,71338,71338,71337,897
R-squared.79.91.80.84
 C. Heterogeneous Effects, Depending on Initial Export Share
Average success rate × oil price.211***.261***.048.117
 (.056)(.063)(.271)(.101)
× 1990s export share−.184***−.128*−.641*−.625***
 (.058)(.075)(.326)(.155)
N38,71338,71338,71337,897
R-squared.79.91.80.84
 D. Heterogeneous Effects, Depending on Initial Skill Intensity
Average success rate × oil price.251***.317***−.300.255*
 (.075)(.080)(.379)(.139)
× 1990s skill intensity−.022***.022**−.185***−.038**
 (.006)(.009)(.038)(.018)
N15,93115,93115,93115,534
R-squared.79.92.80.84
 E. Heterogeneous Effects, Depending on Initial Labor Intensity
Average success rate × oil price.186***.242***−.040.037
 (.054)(.061)(.271)(.110)
× 1990s labor intensity−.004−.025***.003.090***
 (.010)(.009)(.037)(.023)
N38,71338,71338,71337,897
R-squared.79.91.80.84

Note. Panel regressions with firm and industry-year fixed effects. Standard errors in parentheses clustered at kabupaten level.

*p < .10.

**p < .05.

***p < .01.

View Table Image: 1 | 2

Table 2 focuses on wages and unit values. Allcott and Keniston (2017) argued that the route to manufacturing contraction goes through a rise in local manufacturing wages but that prices should be exogenous or not affected by windfalls, as they are set in larger markets. In panel A1, we estimate the average effect across all firms and find a positive and significant effect of oil and gas windfalls on wages. A 10% increase in windfalls leads to a 1.8% increase in wages. Columns 2 and 3 suggest that this rise in wages is driven by wages in production jobs, rather than nonproduction jobs. Column 4 suggests that the firms’ unit values of their core products are unaffected. The latter is consistent with competition, either from firms in other districts or other countries, keeping prices in check. So far, Dutch disease mechanics seem to be at play.14

In panel A2, we control for a time trend interacted with luck, to ensure that our results are not picking up different time trends across districts. The results are unaffected. Panel A3 restricts our sample to firms that existed in 1990, to match the sample in the heterogeneity analysis following in table 2, panels B–E. The results stay the same.

Motivated by Melitz (2003) and the large related literature on heterogeneous firms, we next allow for different effects of the windfalls depending on the firm’s initial labor productivity, export intensity, skill intensity, and labor intensity. We define labor productivity as output per employee, export intensity as the share of output that is exported, skill intensity as the ratio of nonproduction workers to production workers, and labor intensity as total wages over the value of machines.15 We calculate initial levels as the average level in the first three years observed in the 1990s. We start by examining how the effect of windfalls on wages and unit values depends on such firm heterogeneity (table 2). Table A2 in the appendix presents calculations of the effects for the 25th, mean, and 75th percentile of the four dimensions of heterogeneity, to show how the relationships differ for firms placed across the distribution of each characteristic.

In panels B–E of table 2, we find significant differences across these different firms. In column 1 we show that low-productivity, nonexporters, and low skill intensive firms all have significantly higher overall wage growth than other firms. This is driven by both production workers and nonproduction workers in the low-productivity firms and nonexporters, and the nonproduction workers for the firms with low skill intensity. Firms with low labor intensity show this only for production workers. At the 25th percentiles, that is, on the left side of the firm distributions of our four characteristics of heterogeneity, we find that the wages increase between 1.9% and 3.2% in response to a 10% windfall. The numbers for production workers are 2.5%–3.3% and for nonproduction workers 0.5%–2.7%, with the exception of a 0.5% relative fall in the wages of nonproduction workers in firms with high labor intensity (see table A2). We conclude that there is robust evidence that windfalls cause wages to increase.

The results for unit values, presented in column 4, indicate that firms with higher productivity, higher export share, or higher skill intensity increase their unit values significantly less than other firms do. This is consistent with the pattern for wages above. As the overall effect on unit values is insignificant and close to zero, one interpretation is that firms to the right of the productivity distribution, who should be the most able to overcome fixed costs of exporting to other districts or countries, meet prices that are partly determined outside of the district. This keeps their unit values in check, which is an important result. It demonstrates that these firms are subject to Dutch disease forces, that is, wages rising but output prices staying fixed. Yet, they do not succumb to the contraction predicted by theory, as we explore in the next section.

2.2.2.  Margins at Which Firms Adjust

Table 3 looks at 11 more outcomes at the firm level. The structure is the same as before: panel A presents estimates of the mean effect, while panels B, C, D, and E present estimates where we allow for heterogeneous effects according to the four dimensions defined in the previous subsection.

Table 3. 

Margins of Adjustment

 Exit
(1)
Entry
(2)
Product Drop
(3)
Product Intro.
(4)
Output
(5)
Empl.
(6)
Labor Prod.
(7)
% Exported
(8)
Share Imported Inputs
(9)
Intermediate Inputs
(10)
Machines
(11)
 A1. Mean Effect across All Firms
Average success rate × oil price−.021−.005−.013−.032.159**.078**.186*−.039.013.138*−.373
 (.050)(.004)(.033)(.030)(.070)(.030)(.094)(.024)(.011)(.073)(.605)
N49,79549,79549,79549,79549,79449,79549,79349,79549,46949,79248,609
R-squared.33.27.54.55.93.96.80.69.87.92.67
 A2. Mean Effect across All Firms, Controlling for Time Trend Interacted with Success Rate
Average success rate × oil price−.021−.005−.013−.032.158**.078**.186*−.039.013.138*−.371
 (.050)(.004)(.033)(.030)(.070)(.030)(.094)(.024)(.011)(.073)(.606)
Average success rate × time trend.000−.000.000.000−.000.000*−.000.000−.000***−.000.000
 (.000)(.000)(.000)(.000)(.000)(.000)(.000)(.000)(.000)(.000)(.000)
N49,79549,79549,79549,79549,79449,79549,79349,79549,46949,79248,609
R-squared.33.27.54.55.93.96.80.69.87.92.67
 A3. Mean Effect across Firms Existing in 1990
Average success rate × oil price−.002−.004−.008−.031.135*.061*.189*−.047*.008.089−.372
 (.045)(.003)(.032)(.031)(.072)(.032)(.094)(.025)(.010)(.081)(.658)
N38,71338,71338,71338,71338,71238,71338,71138,71338,48238,71038,575
R-squared.32.24.55.56.94.96.80.68.88.92.64
 B. Heterogeneous Effects, Depending on Initial Labor Productivity
Average success rate × oil price.139**.011.117**.0102.155***−.304***2.920***−.028.0761.816***1.274
 (.063)(.009)(.055)(.049)(.183)(.064)(.209)(.054)(.069)(.212)(2.091)
× 1990s labor productivity−.016***−.002*−.014*−.005−.223***.040***−.302***−.002−.008−.191***−.182
 (.004)(.001)(.007)(.006)(.017)(.007)(.021)(.005)(.007)(.021)(.269)
N38,71338,71338,71338,71338,71238,71338,71138,71338,48238,71038,575
R-squared.32.24.55.56.94.96.80.68.88.92.64
 C. Heterogeneous Effects, Depending on Initial Export Share
Average success rate × oil price.006−.003−.013−.033.210***.058*.255**−.026.009.161*−.317
 (.045)(.003)(.031)(.031)(.076)(.033)(.095)(.022)(.011)(.086)(.662)
× 1990s export share−.056**−.009.038*.018−.541***.017−.480***−.150***−.011−.525***−.402
 (.027)(.007)(.022)(.023)(.093)(.036)(.110)(.033)(.023)(.099)(.556)
N38,71338,71338,71338,71338,71238,71338,71138,71338,48238,71038,575
R-squared.32.24.55.56.94.96.80.69.88.92.64
 D. Heterogeneous Effects, Depending on Initial Skill Intensity
Average success rate × oil price−.012.002−.001−.014.252***.072*.286***−.007−.006.169*.396
 (.033)(.004)(.032)(.038)(.088)(.038)(.093)(.023)(.019)(.096)(.934)
× 1990s skill intensity−.004−.000.000.002.002.027***−.016.000.002−.008.182*
 (.003)(.001)(.004)(.004)(.011)(.006)(.012)(.003)(.002)(.013)(.105)
N15,93115,93115,93115,93115,93115,93115,93115,93115,85215,93115,868
R-squared.32.27.57.58.94.96.81.69.87.93.65
 E. Heterogeneous Effects, Depending on Initial Labor Intensity
Average success rate × oil price−.001−.004−.008−.031.138*.060*.192**−.047*.008.092−.371
 (.045)(.003)(.032)(.031)(.072)(.032)(.094)(.025)(.010)(.080)(.654)
× 1990s labor intensity.010**.001−.006−.007.043***−.013**.045***−.002.009*.044***.012
 (.004)(.001)(.004)(.004)(.014)(.005)(.012)(.003)(.004)(.016)(.097)
N38,71338,71338,71338,71338,71238,71338,71138,71338,48238,71038,575
R-squared.32.24.55.56.94.96.80.68.88.92.64

Note. Panel regressions with firm and industry-year fixed effects. Standard errors in parentheses clustered at kabupaten level.

*p < .10.

**p < .05.

***p < .01.

View Table Image: 1 | 2 | 3

Columns 1–4 of panel A in table 3 show no effect on exit rates, entry rates, and the probability of product introduction or product drop.16 This suggests that there is no crowding out of the manufacturing sector due to oil and gas windfalls, in contrast to the standard Dutch disease theory and the symptoms in terms of increased wages observed in the previous subsection.

We observe in columns 5–6, also in contrast to the Dutch disease theory, increases in output and employment due to windfalls. The increase in output is larger than in employment, resulting in an increase in labor productivity, as shown in column 7. We find in column 8 some weak evidence that the share of output exported falls and in column 10 that purchases of intermediate inputs increase, but the coefficients are not statistically significant across all three specifications in panel A.

Panels B–E in table 3 present the heterogeneity results. Table A3 in the appendix presents calculations of the effects at the 25th, mean, and 75th percentile of the heterogeneity dimensions. Panel B shows that low productivity firms are significantly more likely to exit. A firm at the 25th percentile has a 9% likelihood of exit due to an increase in oil and gas windfalls of 10%, compared to the estimated likelihood of 0% for the average firm in the same sample. Such a firm at the 25th percentile is also more likely to drop products. At the same time, surviving firms at the left of the productivity distribution increase output, labor productivity, and use of intermediate inputs significantly more than other firms do. When windfalls increase by 10%, firms at the 25th percentile in terms of initial labor productivity grow their output by 3.5%, labor productivity by 4.7%, and use of intermediate inputs by 2.7%. Employment increases less in the low productive firms than in their more productive peers, however, with a 0.2% versus a 0.9% increase in employment for firms at the 25th and 75th percentile, respectively. These results are indicative of an overall productivity improvement: windfalls induce higher employment expansions in initially high productive firms, while initially less productive firms increase their productivity.

Panel C shows that firms with initial low exports behave similarly to firms with low initial productivity levels. They have significantly higher likelihood of exit and stronger expansions of output, labor productivity, and use of intermediate inputs compared to other firms. The striking difference is in terms of exports, for which export-oriented firms show a significant contraction in the share of their production that is exported. However, the firms at the 75th percentile and the mean in terms of export share also show resilience as they are not contracting in terms of output, employment, or labor productivity (see bottom row of cols. 5–7 in the second panel of table A3). Thus, the lower percentage exported may just reflect a boom in the home market instead of a pulling back from international markets.

The results according to initial skill intensity are presented in panel D. Firms with high skill intensity stick out only by significantly higher increases in employment and machine investment. A 10% windfall increase induces a firm at the 75th percentile in terms of skill intensity to increase employment by 0.7% and increase the value of machines by 3.7%.

Finally, panel E reveals that initially more labor intensive firms are more likely to exit in response to the windfalls, with a firm at the 75th percentile of initial labor intensity having about the same exit probability as a firm at the 25th percentile of initial labor productivity (table A3 in the appendix). Compared to other firms, surviving labor intensive firms expand output more and increase employment less, resulting in a higher increase in labor productivity. The initially labor intensive firms also expand their use of intermediate inputs, both imported and in general, more than other firms. Recall also that windfalls led to higher unit values for firms with initially high labor intensity in table 3. The combination of higher prices and expanded output is consistent with a positive demand shock facing these initially labor intensive firms.

The Dutch disease resistance, and even output expansions, we observe for Indonesian manufacturing firms on average is a composite of several adjustments. Tables A2 and A3 in the appendix summarize the effects. First, firms initially characterized by either low productivity, nonexporting or high skill intensity, that is, the “bad” ends of these distributions, are more likely to exit when faced with a windfall.17 At the same time, output, labor productivity, and the use of intermediate inputs grow more at the “bad” end of all four distributions. There is a die or thrive bifurcation for firms located at the “bad” ends of our dimensions of heterogeneity. At the “good” end, output and labor productivity typically expand less, but employment either increases or keeps up. This is consistent with “good” firms having fewer low-hanging fruit left to pick in terms of productivity improvements.

2.2.3.  Summing Up the Empirical Evidence on Heterogeneity

In sections 2.2.12.2.2 we document important heterogeneity across firms in the adjustments to oil and gas windfalls. Figure 6 illustrates for three outcomes how the estimated effects of windfalls vary across the distributions of firms in terms of initial labor productivity, export share, skill intensity, and labor intensity. For most of the firms, windfalls have led to a significant increase in wages. This effect is highest for low skill intensive firms and nonexporters. Yet it is only for low-productivity firms that this translates into a positive probability to exit the market. Low productivity firms, on the other hand, along with nonexporters and labor intensive firms, also expand their output the most. These results are consistent with firm expansions due to the windfalls, but with some firms going out of business.

Figure 6. 
Figure 6. 

Heterogeneous effects, based on regression in table 3. The solid dark line plots the effect of windfalls on the outcome variable, written on top of the graph, the dashed lines are 95% confidence intervals, and the inverted U-shaped line gives the distribution of labor productivity, labor intensity, or share of output exported across firms in 1990s.

The estimates in section 2.2.12.2.2 are based on regressions with firm and sector-year fixed effects and therefore do not reflect reallocations across firms or sector-specific time shocks. To get a sense of the effects across different sectors, we plot in figure 7 the coefficients on the windfall measure from sector-specific regressions with firm and year fixed effects included. On the horizontal axis we measure the coefficients when we use wages as the dependent variable, and on the vertical axis the coefficients when we use the exit probability as the dependent variable. Most of the sectors see a zero or positive effect of windfalls on wages. For the exit probability, sectors making products such as leather products, furniture, and wood experience increased exit probabilities, whereas most of the others experience zero or negative effects. Zooming out to look for a correlation between the effect of wages and the effect on exit rates in figure 7, there seems to be no obvious link between the two at the sector level. Instead, this may suggest that it is within-sector heterogeneity that affects firm exit, whereby higher-performing firms are better able to adapt and survive.

Figure 7. 
Figure 7. 

Responses in wages and exit probability, sector by sector. The graph shows coefficients from our baseline regression run industry by industry.

3.  Conclusions

We document Dutch disease resistance in a large developing country, Indonesia. Manufacturing firms in booming oil and gas districts show Dutch disease symptoms in that the boom leads them to pay higher wages, while they on average are unable to respond by raising the unit values of their products. In contrast to the Dutch disease hypothesis, however, we find no evidence to suggest that firms on average are more likely to exit in booming districts. Instead, firms raise their labor productivity and expand their employment and output. Our results are thus in line with previous evidence from the United States showing that the local manufacturing sector may benefit from, rather than getting crowded out by, local resource booms. This study further highlights the mechanisms of the Dutch disease resistance by investigating a host of outcomes at the firm level and by allowing for firm heterogeneity within sectors.

First we find that the firms with initially higher productivity, higher export share, or higher skill intensity increase wages and their unit values less than other firms do. These firms may be the ones active in markets outside of their local district as they overcome the fixed costs of exporting and experience that their output prices de facto are fixed. Of particular interest may be the firms engaged in international exports, of which many expand their output, employment, and labor productivity due to the boom. They show a significant contraction in the percentage of their production that is exported, which may at least partly reflect a boom in the home market instead of a pulling back from international markets. Firms with very high initial export shares show evidence of a contraction in output.

Second, our results point to increased labor productivity for the manufacturing sectors through three different channels: exit of initially low productivity firms, strong productivity (as well as output and employment) increases among surviving initially low-productivity firms, and strong employment increases in initially high-productivity firms. These findings are consistent with low-productivity firms being vulnerable to the increase in wage costs, surviving low-productivity firms having much room for productivity improvements and high-productivity firms taking on more workers but having less room for improving their productivity.

Third, we find that initially more labor intensive firms do surprisingly well, except for the fact that they are more likely to exit in response to the windfalls, just like the firms with low initial labor productivity. Compared to initially more capital intensive firms, the initially labor intensive firms see larger increases in their labor productivity, through larger output expansions and lower employment expansions, and an expanding use of intermediate inputs.

These findings provide hypotheses for future research. The current results in the literature suggest that manufacturing firms in booming districts are able to expand with the boom, and they may be able to do so by reorienting their production toward the local market. Future research could investigate how the market orientation of manufacturing firms changes due to local resource booms, in terms of reallocations across geographic markets (within districts, across districts within country, and across countries) as well as in terms of reallocations across different product markets. Future research could also study the productivity effects in more detail. To what extent does the more expensive labor lead to more capital intensive production versus technology upgrading? What is the role of reallocation across firms versus within-firm adjustments? The use of firm-level data to study the consequences of local resource booms has suggested that local resource booms are a blessing rather than a curse also for manufacturing and places some doubt on the microfoundations for the Dutch disease hypothesis.

Appendix

Table A1. 

Summary Statistics

 NMeanSDP25P75MinMax
Wage48,1949.09.998.529.72016.38
Wage production48,19413.191.7611.9514.37021.93
Wage nonproduction48,194105.039.3113.38021.03
Unit value47,0633.832.91.715.63020.47
Exit48,194.05.210001
Entry48,194.01.080001
Product drop48,194.42.490101
Product intro.48,194.4.490101
Output48,19315.662.3313.8317.388.0824.5
Employment48,1945.091.274.035.893.6910.94
Labor prod.48,19310.181.499.1511.071.4717.55
% exported48,19411.6729.7000100
Share imported inputs47,899.12.280001
Intermediate inputs48,19215.082.4413.216.87024.47
Oil price (2010 USD)48,19434.5912.3730.144.1717.0160.87
Average success rate48,194.76.3.6101
Average success rate × oil price48,1942.651.072.053.4104.11
Labor productivity in 1990s5,6909.061.348.109.864.6714.6
Labor intensity in 1990s5,690−.0931.80−1.321.07−10.99.20
Export share in 1990s5,690.13.270.08301
Skill intensity in 1990s2,286−1.482.75−3.06−.14−8.939.65

Note. Wage, unit value, output, employment, input, and productivity variables are in inverse hyperbolic sines. The oil price is logged when interacted with the average success rate.

View Table Image
Table A2. 

Calculated Heterogeneous Effects on Wages and Unit Values

 Value of Characteristic
(0)
Wage
(1)
Wage Production
(2)
Wage Nonproduction
(3)
Unit Value
(4)
WF 1.4461.059.8691.592
WF × labor productivity −.139−.09−.101−.173
P258.100.320.330.051.191
Mean9.060.187.244−.046.025
P759.860.075.172−.127−.114
WF .211.261.048.117
WF × export share −.184−.128−.641−.625
P25.000.211.261.048.117
Mean.130.187.244−.035.036
P75.083.196.250−.005.065
WF .251.317−.300.255
WF × skill intensity −.022.022−.185−.038
P25−3.060.318.250.266.371
Mean−1.480.284.284−.026.311
P75−.140.254.314−.274.260
WF .186.242−.04.037
WF × labor intensity −.004−.025.003.09
P25−1.320.191.275−.044−.082
Mean−.093.186.244−.040.029
P751.070.182.215−.037.133

Note. WF (windfalls) refers to “Average success rate × oil price.” Labor productivity, export share, skill intensity, and labor intensity are all measured in the 1990s, as in the rest of the paper. Value of characteristic refers to the 25th, mean, or 75th percentile of the measures of labor productivity, export share, skill intensity, and labor intensity, respectively, as presented in table A1.

View Table Image
Table A3. 

Calculated Heterogeneous Effects on the 11 Margins

 Value of Char.
(0)
Exit
(1)
Entry
(2)
Product Drop
(3)
Product Intro.
(4)
Output
(5)
Empl.
(6)
Labor Prod.
(7)
% Exported
(8)
Share Imported Inputs
(9)
Intermediate Inputs
(10)
Machines
(11)
WF .139.011.117.0102.155−.3042.920−.028.0761.8161.274
WF × labor productivity −.016−.002−.014−.005−.223.040−.302−.002−.008−.191−.180
P258.100.009−.005.004−.031.349.020.474−.044.011.269−.184
Mean9.060−.006−.007−.010−.035.135.058.184−.046.004.086−.357
P759.860−.019−.009−.021−.039−.044.090−.058−.048−.003−.067−.501
WF .006−.003−.013−.033.210.058.255−.026.009.161−.317
WF × export share −.056−.009.038.018−.541.017−.480−.150−.011−.525−.402
P25.000.006−.003−.013−.033.210.058.255−.026.009.161−.317
Mean.130−.001−.004−.008−.031.140.060.193−.046.008.093−.369
P75.083.001−.004−.010−.032.165.059.215−.038.008.117−.350
WF −.012.002−.001−.014.252.072.286−.007−.006.169.396
WF × skill intensity −.004.000.000.002.002.027−.016.000.002−.008.182
P25−3.060.000.002−.001−.020.246−.011.335−.007−.012.193−.161
Mean−1.480−.006.002−.001−.017.249.032.310−.007−.009.181.127
P75−.140−.011.002−.001−.014.252.068.288−.007−.006.170.371
WF −.001−.004−.008−.031.138.060.192−.047.008.092−.371
WF × labor intensity .010.001−.006−.007.043−.013.045−.002.009.044.012
P25−1.320−.014−.005.000−.022.081.077.133−.044−.004.034−.387
Mean−.093−.002−.004−.007−.030.134.061.188−.047.007.088−.372
P751.070.010−.003−.014−.038.184.046.240−.049.018.139−.358

Note. WF (windfalls) refers to “Average success rate × oil price.” Labor productivity, export share, skill intensity, and labor intensity are all measured in the 1990s, as in the rest of the paper. Value of characteristic refers to the 25th, mean, or 75th percentile of these measures; see table A1.

View Table Image

Notes

Due to a production error, this article was initially published with an incorrect page range, pp. 1019–1051, and was reposted on October 22, 2019, with the correct pagination. The publisher regrets the error.

James Cust is at the World Bank and OxCarre (). Torfinn Harding is at the NHH Norwegian School of Economics and Oxcarre (). Pierre-Louis Vézina (corresponding author) is at King’s College London and OxCarre (). We would like to thank two anonymous referees, Ryan Edwards, Beata Javorcik, Peter Neary, and Anthony Venables as well seminar participants at OxCarre, King’s, NHH, and the 2016 Royal Economic Society meeting for helpful comments. We are grateful to Wood Mackenzie for academic access to their PathFinder database. Support from the BP-funded Oxford Centre for the Analysis of Resource Rich Economies (Oxcarre) and the Equinor chair in Economics at NHH is gratefully acknowledged. The findings, interpretations, and conclusions do not necessarily reflect the views of the World Bank, its executive directors, or the governments they represent.

1. See Bryan and Morten (2019) for a study of the effects of migration frictions in Indonesia. They show that a 10% reduction in the distance between two locations leads to a 7% increase in the proportion of migrants who flow between the two locations, thus highlighting the presence of movement costs.

2. Harding et al. (2016) document that giant oil and gas discoveries appreciate the real exchange rate through the prices of nontradable goods whereas the prices of traded goods are not affected.

3. Allcott and Keniston (2017) find resilience of manufacturing in US firm-level data, and the results of Cust and Rusli (2014) and Cavalcanti et al. (2019) of high growth in resource-rich Brazilian municipalities and Indonesian districts, respectively, also go in the same direction. Jacobsen and Parker (2016) find that resource booms in the American West did not affect employment in manufacturing. Like us, Benguria et al. (2018) find that windfalls increase wages, especially for unskilled workers in Brazil. In the context of mining, De Haas and Poelhekke (2019) find that local mining activities negatively affect the business environment and the growth of firms in the tradable sector, whereas firms in the nontradable sector are positively affected.

4. At the macrolevel, Bjørnland and Thorsrud (2016) find substantial positive productivity spillovers from the resource sector to nonresource sectors for Australia and Norway. Bjørnland et al. (2018) formalize such effects in a theory model, where they argue that the “resource movement effect” in Corden and Neary (1982) may have positive productivity effects on the manufacturing sector, in contrast to the “spending effect.”

5. In emerging economies, governments might also have less capacity to enact policies to mitigate Dutch disease effects, and there might be more trade and migration frictions that could push the effect in one direction or the other. We might also expect different effects due to more surplus labor and lower quality of the infrastructure.

6. Moshiri et al. (2019) consider firm heterogeneity in terms of productivity. Other studies on manufacturing have used sector-level data on value added and employment or sector-level data on trade. For example, Smith (2014) examined the impact of the oil price boom in the 1970s and the subsequent bust on manufacturing in oil-dependent countries. He showed that manufacturing exports, value added, wages, and employment actually increased during the boom and that these effects decreased during the bust, albeit gradually. Ismail (2010) used manufacturing sector data across countries and found evidence of a negative impact of oil prices on manufacturing output, consistent with the Dutch disease. Harding and Venables (2016) study sector-level trade data and find contractions in nonresource exports. Arezki, Fetzer, and Pisch (2017) examine the comparative advantage effects of the US fracking boom on US manufacturing exports by sector classification. They find evidence supporting that relatively cheaper US gas has boosted US manufacturing in terms of exports, while on the intensive margin new energy intensive manufacturing has been added as measured by large new investments. In terms of studies not focusing on manufacturing, James and Aadland (2011) showed that resource-dependent US counties exhibit more anemic economic growth, Cavalcanti et al. (2019) showed that oil discoveries significantly increase per capita GDP and urbanization across Brazilian municipalities, and Aragon and Rud (2013) found positive effects of a large gold mine on real income in Peru. Other studies include Brollo et al. (2013), Caselli and Michaels (2013), and Borge et al. (2015), who use variation across municipalities to study the response of local politicians to resource windfalls. The two former find support in favor of the political resource curse in Brazilian municipalities, while the latter finds that higher local government revenue from hydropower reduces the efficiency in production of public goods across Norway. Edwards (2015) examines the welfare effects of the palm oil boom at the subnational level in Indonesia. For an overview of the literature on within-country effects of natural resources, see Cust and Poelhekke (2015). Other earlier empirical work on the Dutch disease includes Chen and Rogoff (2003) and Cashin et al. (2004) on the real effective exchange rate.

7. Other strategies used in the literature are to use initial or early values of the weights across districts, like Bazzi and Blattman (2014), or instrumenting the weights, like Dube and Vargas (2013). Our strategy is similar to that used by Cotet and Tsui (2013), who also exploited randomness in the success or failure of oil exploration to look at the effect of oil wealth on conflict across countries. It is also similar to Cavalcanti et al. (2019), who suggest that districts with drilling but no discovery are a valid counterfactual to those who make a discovery and go on to produce oil. Cassidy (2019) utilizes information on geological basins to create an instrument for oil production across countries. Allcott and Keniston (2017) use as weights a measure of the initial economically recoverable oil and gas endowment for each county. Their measure include subsequent production, proven reserves, and undiscovered reserves. Instead of prices, they use time series variation in national oil and gas employment.

8. The approach of interacting district shares with country-wide time variation is close in spirit to the approach of Bartik (1991). Goldsmith-Pinkham et al. (2018, 5) find that for Bartik-like instruments, which they define as “[an instrument] that uses the inner product structure of the endogenous variable to construct an instrument,” the identifying assumption relates to the shares. They quote Baum-Snow and Ferreira (2015, 50): “validity [of the Bartik instrument] … relies on the assertion that neither industry composition nor unobserved variables correlated with it directly predict the outcomes of interest conditional on controls.” Brunnschweiler and Bulte (2008), van der Ploeg and Poelhekke (2010), and Cust and Harding (2019) highlight the issue of potentially endogenous resource measures in empirical work on the resource curse. A recent contribution by Smith and Wills (2015) uses a combination of oil prices and discoveries to show that resources increase GDP nationally. Recently, several papers have studied the effects of giant oil and gas discoveries on outcomes such as conflict (Lei and Michaels 2014), macroeconomic adjustments (Arezki, Ramey, and Sheng 2017), and the real exchange rate (Harding et al. 2016). With our detailed drilling data we are able to construct discovery rates, which is the closest one can get to observe the actual geology. Although seismic data on geology are available for many places, the uncertainty about the geology is surprisingly high. The average discovery rate conditional on drilling is about 60% globally, with large variation by location.

9. We define dry wells (vs. those recorded as a discovery) using the industry standard definition. A dry well or dry hole in our data is denoted as one that is evaluated to contain insufficient oil for commercial production. Wood Mackenzie analysis supplements the raw well data to include postdrilling well evaluation, allowing for a more accurate picture of the eventual result and assessment of the well, and not just the recorded result at the time of drilling.

10. Cotet and Tsui (2013) also exploit randomness in success of oil exploration to look at the effect of oil wealth on conflict. They find that oil discoveries do not increase the likelihood of violent challenges to the state.

11. We use constant 1998 district definitions in our analysis, linking firms to origin districts, rather than new districts that emerge from subsequent splits later in the period.

12. We examined the relationship between luck and district-level characteristics: initial GDP, initial population, district-level institutional quality, city dummy, built-up environment dummy, road density, and travel time to market. We find no statistically significant relationship between these characteristics.

13. Since the data only cover plants with more than 20 employees, our exit rate is measured with error. We treat firms that shrink below 20 employees as exiting, and firms that hire just above the threshold appear as entering. This suggests that our exit and entry rates may be slightly overestimated. This should not affect our estimates, however, as luck in discovery should not be correlated with firm size. If anything, the data suggest that firms may be slightly larger in lucky districts; hence we could be underestimating the effect of windfalls on exit rates. Since we find exit rates of the same magnitude in districts with and without oil production, as well as in lucky and unlucky districts, we do not worry too much about this measurement error. We also find that windfalls lead to increases in employment on average, suggesting that we are not likely to be underestimating the effect on the exit rate.

14. Average success rate is the average of the discovery rate, or luck as defined in our empirical strategy. Production and nonproduction wage are for white-collar vs. blue-collar workers, labeled as production/nonproduction workers in the data. Unit value is (value/quantity) of the firm’s core product, defined as the one with maximum value.

15. We define skill intensity as ln(Number of Nonproduction Workers/Number of Production Workers) and labor intensity as ln(Wages/Machines).

16. Note that we always have the same firms in the sample, and some of these enter or exit during our sample period. We define exit as a dummy equal to 1 if it is the firm’s last year, and entry if it is the firm’s first year.

17. By “good” we refer to high initial values of labor productivity, export share, and skill intensity and low initial values of labor intensity.

References