Policy Research Working Paper 8415 Jobs! Electricity Shortages and Unemployment in Africa Justice Tei Mensah Africa Region Office of the Chief Economist April 2018 Policy Research Working Paper 8415 Abstract To what extent does unreliable electricity provision perva- approaches, the paper shows that outages have a non-triv- sive in many African countries affect job creation in the ial negative impact on employment. The effect is driven region? This paper addresses the question by assembling by a reduction in employment in non-agricultural sectors household and firm level data from 29 African countries and skilled jobs. Unskilled jobs are unaffected by electricity along with unique project level data on foreign direct outages. The negative effect of outages on firm entry and investment (FDI). Leveraging several quasi-experimental the performance of incumbent firms are plausible channels.. This paper is a product of the Office of the Chief Economist, Africa Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The author may be contacted at jmensah2@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Jobs! Electricity Shortages and Unemployment in Africa Justice Tei Mensah∗ Originally published in the Policy Research Working Paper Series on April 2018. This version is updated on April 2023. To obtain the originally published version, please email prwp@worldbank.org. JEL: E24, F14, L94, O10, O15, O55 ∗ Office of the Chief Economist, Africa Region, The World Bank. Email:jmensah2@worldbank.org. This article has benefited from comments and discussions with Pete Klenow, Moussa Blimpo, Morgan Hardy, Wilfried Kouam´ e, Franklin Amuakwa-Mensah, Th´ eophile Bougna, Andrew Agyei-Holmes, Aimable Nsabi- mana, Anthony Amoah, Abenezer Aklilu, Francis Annan, Kibrom Abay, Vivien Foster, Belinda Archibong, Sabah Abdulla, and Valentina Saltane. Comments from seminar participants at: InfraXchange at the World Bank (2021), CSAE Conference on African Development (2018), Annual Bank Conference on Africa (2018), African Development Bank (2018), University of Ghana (2018), Kwame Nkrumah University of Science and Technology (2018), and the WGAPE Workshop at UCLA (2018) are also acknowledged. The findings, interpretations, and conclusions expressed in this paper are entirely those of the author. They do not neces- sarily represent the views of The World Bank Group and its affiliated organizations or those of the Executive Directors of The World Bank Group or the governments they represent. 1 Introduction Electricity is considered one of the basic attributes of modern life. It is a key input for the production of goods and services, as well as quality of life. The reality, however, is that despite the fact that more than 580 million people in Africa lack access to electricity (IEA, 2019), the quality of supply to connected households and firms is precarious. Elec- tricity outages have become a common feature in many African countries (Andersen and Dalgaard, 2013; Blimpo and Cosgrove-Davies, 2019). A large body of development literature has underscored the importance of electricity access on socioeconomic outcomes such as education, income, health and labor allocation (Dinkelman, 2011; Lipscomb et al., 2013; Abbasi et al., 2022). Yet little is known about the economic impact of unreliable electricity services. Available studies on the impact of elec- tricity outages have largely focused on the extent to which outages affect firm productivity (Allcott et al., 2016; Abeberese et al., 2021) and profitability (Cole et al., 2018; Hardy and McCasland, 2019). An important yet often ignored question is the extent to which persis- tent electricity shortages affect job creation and consequently, the rate of unemployment in the developing world. The main goal of this paper is to show evidence of how electricity shortages1 constrain job creation in the developing world. Specifically, using instrumental variable regression, difference-in-difference, and fixed effect estimators with recent data on households (in- dividuals) and firms in 29 African countries, I estimate the causal impact of electricity shortages on employment in Africa, and document the mechanisms through which the supply inefficiencies affect job creation. The paper hypothesizes and tests two main chan- nels through which persistent electricity outages affect job creation and hence unemploy- ment: (i) on the extensive margin, persistent outages create distortions in the business climate and increase the expected cost of doing business. This can discourage potential en- 1 In the remainder of this paper, electricity outages and shortages are used interchangeably. 1 trepreneurs (investors) from establishing (investing in) businesses that would otherwise have employed people. As a result, persistent outages could reduce entry of domestic and foreign (via foreign direct investment (FDI)) firms;2 (ii) In the intensive margin, shortages in electricity supply exert adverse impact on firms’ productivity and profit, given that elec- tricity is an important factor of production. Therefore the negative impact of outages on firm performance can have negative consequences on firms’ demand for labor. Causal estimation of the impact of infrastructure services such as the quality of elec- tricity is often beset with the challenge of endogeneity, as the incidence and intensity of electricity outages are non-random across space and time.3 Local economic, social and po- litical factors may confound the relationship between outages and the outcome variables of interest. To overcome this challenge of identification, the paper uses several estimation strategies namely: instrumental variable (IV), difference-in-difference (DID), synthetic control, and panel fixed effects. The empirical strategy of the paper is summarized as follows: In the main analysis, I use an IV strategy that exploits plausibly exogenous variations in the incidence of out- ages induced by variations in lightning strikes across space and time. Lightning strikes are known to be a major cause of surges in electrical systems leading to over-voltage and destruction of power infrastructure thereby causing outages (Andersen et al., 2011, 2012; Andersen and Dalgaard, 2013). Leveraging this relationship between lightning activities and electricity outages, I combine granular data on measures of lightning intensity with two rounds of individual surveys from the Afrobarometer dataset in 25 countries to esti- mate an IV regression of the effect of outages on employment. The identifying assumption advanced here is that lightning strikes influence labor market outcomes only through the effect on the quality of electricity supply, i.e, the so-called “exclusion restriction assump- 2 The high cost of business associated with outages could also facilitate the exit of firms either through relocation to other countries (cities) with much reliable electricity supply, or firm shut down. 3 Like most essential services, random assignment of outages across locations is not feasible from a policy perspective and more importantly, unethical. 2 tion”. The main threat to this assumption relates to the possibility of lightning strikes influencing labor market via channels such as information technology (IT) adoption by firms (households) and the consequent effect on demand (supply) for labor as lightning has been shown to affect the diffusion of technologies such as mobile phones and comput- ers (Andersen et al., 2012; Manacorda and Tesei, 2020; Guriev et al., 2020). To address this concern, I control for the diffusion of mobile phone networks (2G, 3G, & 4G) as a proxy for (general) technology adoption. Thus, by partialling out the effect of lightning strikes on the diffusion of technology, the IV strategy exploits variations in outages induced by lightning strikes. In other words, while the exclusion restriction assumption may not hold unconditionally, the assumption is highly plausible conditional on controls such as mobile network penetration, and spatial and time fixed effects. In addition to the cross-country analysis, I conduct two country case studies. First, I exploit a unique quasi-natural experiment in Ghana induced by a four year nationwide power (“Dumsor”) crisis between 2013 and 2016.4 The crisis led to severe power rationing in the country. I estimate the impact of the power crisis on employment by exploiting plausibly exogenous variation in exposure to the crisis using a difference-in-difference design. In addition, I present evidence from Nigeria where I rely on household panel data on employment outcomes and quality of electricity supply to estimate the effect of unreliable electricity supply on employment rates using a panel fixed effect design.5 Finally, in terms of causal mechanisms, I provide evidence on the intensive and exten- sive margins. On the intensive margin, I use firm-level data from 10 African countries to evaluate the effect of outages on firm performance and labor demand using the same IV design. On the extensive margin, I evaluate the role of quality of electricity provision on firm entry (and exit) using two approaches. First, I use firm census data from Ethiopia to show how reliability of electricity influences (net) entry of firms using a fixed effect 4 See https://www.theguardian.com/world/2015/may/17/ghanas-celebrities-lead-protest-marches-against-ongoing-ene and https://en.wikipedia.org/wiki/Dumsor 5 Identification here relies on within household variations in exposure to outages. 3 estimator. Second, I leverage the “Dumsor” power crisis in Ghana to estimate the effect of unreliable electricity provision on entry of foreign firms using greenfield foreign direct investment (FDI) as a proxy. Specifically, I utilize unique data on greenfield FDI projects from fDiMarkets6 and the synthetic control method to estimate the effect of the crisis on FDI in the non-energy-and-construction sectors7 of the Ghanaian economy. The main finding of the paper is that electricity outages exert a non-trivial negative impact on employment. From the cross-country analysis, I find that outages reduce em- ployment by about 13.5 percentage points (pp). The results of the country case studies in Ghana and Nigeria also show a negative effect of outages on employment that are econom- ically and statistically significant, albeit with relatively low magnitudes: the DID estimates from Ghana suggest that the “Dumsor” power crisis increased unemployment by 4.7 pp, while the fixed effects estimates from Nigeria suggest that outages are associated with a 5.7 pp increase in unemployment. Thus, overall the estimates suggest that outages are associated with a 4.7 pp to 13.5 pp increase in unemployment in the region. Additionally, evidence from the paper suggests that the effects are largely concentrated in employment in non-agricultural sectors and skilled jobs. Employment of unskilled workers are unaf- fected by outages. The null effect of outages on employment of unskilled workers provides support to the identification strategy in estimating the causal impact of outages on employ- ment as, in practice, we do not expect unskilled tasks or jobs reliant on manual labor to be affected by outages. I also find suggestive evidence that the job losses associated with unreliable electricity provision are largely concentrated in the private sector. Interestingly, employment in the public sector increases with unreliable electricity provision, albeit the level of increase is relatively lower than the job losses in the private sector, hence the overall reduction in employment. 6 a subsidiary of the Financial Times. See https://www.fdimarkets.com/ 7 The exclusion of FDI to the energy and construction sectors is motivated by two main factors: (i) FDI into the energy sector may increase in direct response to the power crisis, thereby leading to reverse causality; (ii) FDI to the construction sector in Africa is mainly concentrated in the real estate sub-sector which is less reliant on energy. Hence, the power crisis is less likely to affect FDI to the sector. 4 On potential mechanisms through which electricity shortages affect employment, I document two key findings. First electricity shortages reduce the entry of new firms through a reduction firm density and FDI. Evidence from the Ethiopian firm census data shows that areas with high prevalence of outages have lower number of manufacturing firms operating. Incumbent firms also operate for lower durations during the year, as outages force them to operate below optimal capacity and sometimes shut down produc- tion plants during periods of outages. Further, results from a synthetic control method (SCM) estimation of the effect of the “Dumsor” power crisis in Ghana suggest that be- tween 2013 and 2016 (during the crisis), the number of FDI projects to the non-energy- and-construction sectors in the country declined by about 12.3% per annum. High cost of doing business and the unfavorable macroeconomic shocks induced by the crisis are possible reasons for the slump in FDI. As a result, businesses that would have otherwise create jobs were lost. Secondly, the paper shows that outages also affect the performance of incumbent firms. The results show a negative effect of outages on firm revenue and productivity: for every percent increase in the frequency of outages experienced by firms, sales, sales per worker and value-added per worker decline by 1.2%, 1.3% and 2.3% respectively. Further, I find a negative effect of outages on labor demand, particularly, temporary workers. A percent increase in outage frequency (duration) is associated with a 0.58% (0.32%) reduction in the number of temporary workers hired by firms. The effect on demand for full time workers is also negative albeit statistically insignificant. In addition, the results show a negative and significant effect of outages on total labor cost and labor cost per worker. This provides suggestive evidence that firms perhaps respond to the declining revenue as a result of outages by reducing wages. Overall, the findings of the paper suggest that the negative of outages on firm entry and performance of incumbent firms are plausible channels through which outages affect employment. This paper offers two main contributions to the literature. First, to the best of my knowl- 5 edge, this paper presents the first causal evidence on how the provision of unreliable elec- tricity services in the developing world contributes to unemployment in the region. Un- like the existing studies (see for example Allcott et al., 2016; Chakravorty et al., 2014; Alam, 2013; Steinbuks and Foster, 2010; Fisher-Vanden et al., 2015; Reinikka and Svensson, 2002), this paper moves beyond quantification of the level of impact of electricity shortages on firm productivity, to examine the implications on employment. Allcott et al. (2016), for instance, offer evidence on the effects of electricity shortages on the performance of Indian firms. They show that electricity shortages reduce firm revenue by 5 to 10 percent albeit the productivity losses are marginal. In China, evidence from Fisher-Vanden et al. (2015) indicate that firms respond to electricity shortages by re-optimizing input use. Notably, the paper reveals that in spite of the high cost of outsourcing, Chinese manufacturing firms outsource their production in order to mitigate the high productivity losses associ- ated with outages. In the African context, Cole et al. (2018) and Abeberese et al. (2021) also provide evidence of a negative effect of outages on firm performance. What is absent so far, in the literature, is the labor market implications of the impact of outages on the industrial sector. The results from this paper therefore contribute significantly to filling this gap in the literature. Secondly, this paper brings new knowledge to the strand of the literature on the im- pact of distortions in business climate on firms (Restuccia and Rogerson, 2008; Hsieh and Klenow, 2009; Garicano et al., 2016) by showing the implications of the distortions induced by outages on entry of new firms. Restuccia and Rogerson (2008) for instance assert that distortions in the business environment affect productivity and resource allocation. As a result, efficient firms tend to produce too little and employ few workers. These distor- tions have also been shown to account for the productivity gaps between the advanced and developing economies (Hsieh and Klenow, 2009). Evidence from Abeberese (2017) also indicate that high energy cost constrains the ability and incentives of firms to move into high productive energy intensive industries. Apart from providing evidence on the 6 effects of outages on performance of incumbent firms, findings from the paper indicate that outages affect the entry of new firms as well by reducing the incentives of investors to invest in markets riddled with unreliable power supply. The remainder of this paper is structured as follows. Section 2 presents the theoretical underpinnings on the effects of electrification on employment and job creation. Details on data and construction of key variables are presented in Section 3. The identification strategy and results of the paper are outlined and discussed in Section 4. I explore potential mechanisms in Section 5. The paper concludes in Section 6 with a summary of the key findings and implications for policy. 2 Conceptual Framework Technology shocks such as electrification affect economic outcomes in diverse ways. First, electricity affects the nature of home production (Lewis, 2014). Through appliance use, electricity increases labor productivity in home production such as cooking, washing, ironing, etc., thereby reducing total time spent on home activities and freeing up labor for participation in the labor market (Greenwood et al., 2005; Ramey and Francis, 2009; Coen- Pirani et al., 2010; Dinkelman, 2011; Lewis, 2014; Akpandjar and Kitchens, 2017). It also creates an endowment effect through the demand for market goods (iron, fans, fridges, etc) whose utilization has been made possible by the presence of electricity (Dinkelman, 2011). The need for income by households to ‘effectively’ demand these market goods pushes them to supply more labor into the market (Dinkelman, 2011; Lewis, 2014). Electrification also improves the productivity of local economies. Extending electric- ity services to communities enables the adoption and utilization of modern technology ¸a (such as irrigation) to improve labor productivity (Assunc ˜ o et al., 2014; Lewis and Sev- ernini, 2014). For instance, electrification in farming communities can increase mechaniza- ¸a tion and enhance irrigation schemes to improve agricultural productivity (Assunc ˜ o et al., 7 2014; Lewis and Severnini, 2014). Access to electricity can also spur technology adoption among local (cottage) industries thereby boosting productivity and possible spillover ef- fects on employment and wages (Fried and Lagakos, 2021). Further, electrification like most technology shocks offer opportunities for the creation of new businesses and induce structural change (Fried and Lagakos, 2021). Access to electricity can foster the creation of new jobs as it spurs prospective entrepreneurs to take advantage of the enabling conditions the infrastructure provision offers. For instance in many developing economies where the informal sector dominates, access to electricity can enable households to set up small firms that produce intermediate or final goods (and services) for the market. Additionally, appliance use and modernization of agricul- ture and local industries through the use of electricity can shift employment into skilled non-agricultural employment, thereby reducing the share of labor employed in agricul- ture. A booming non-agricultural sector increases demand for labor and thus reduces out-migration in electrified communities, while attracting labor from neighboring local- ities to participate in the booming economy (Fried and Lagakos, 2021). Implicit in the above is the assumption that electricity services are stable and reliable. Meanwhile, elec- tricity outages are pervasive in many developing countries thereby constraining the real- ization of the full impact of electrification. In this section, I outline three channels through which unreliable electricity provision affect job creation. First, the quality of electricity supply can affect firm entry and exit. Persistent out- ages signal high production cost and uncertainties in the business climate thereby reduc- ing the incentive(s) of potential entrepreneurs (investors) in establishing (investing in) businesses. For instance, persistent outages may reduce foreign direct investment in non- energy intensive sectors such as manufacturing due to the associated effects of unreliable power supply on cost of doing business.8 A negative effect of outages on investments into 8 Investment in the energy sector may however increase as the outages may be associated with factors such as capacity constraints or underdevelopment of the power sector, hence the potential for higher returns on investment(s) in the sector. 8 greenfield (new) projects9 could reduce firm entry and hence job creation. In addition, firms may respond to unreliable electricity provision by relocating to regions (countries) with reliable access to electricity or shut down production to avoid investment losses.10 Thus, pervasive outages constrain expansion of the industrial and service sectors with direct and indirect impacts on job creation. The second channel relates to the effect on firm performance. A plethora of evidence suggest that electricity shortages impose significant losses in productivity and profitabil- ity (Fisher-Vanden et al., 2015; Allcott et al., 2016; Cole et al., 2018). These impacts have labor market implications: firms respond to these adverse productivity shocks by reduc- ing variable cost through job cuts or reducing wages. In addition, some firms respond to electricity supply uncertainties by either substituting materials for energy inputs or out- sourcing intermediate production to external firms (Fisher-Vanden et al., 2015). These strategies, particularly, outsourcing to external firms often result in layoffs. Finally, outages can affect trade and export competitiveness of firms. Unplanned out- ages distort production schedules of firms reliant on grid supply of electricity thereby af- fecting their ability to adequately meet the needs of their (domestic/international) clients. Firms reliant on in-house electricity generation also face high energy cost, due to the high cost per kWh of in-house generation relative to grid supply (Steinbuks and Foster, 2010). High energy cost increases production cost and output prices thereby affecting the com- petitiveness of firms on export markets. Hence persistent outages could have negative implications on employment in a country with a buoyant export sector as exporting firms may struggle to survive. In this paper, I provide evidence on the effects of electricity outages on firm perfor- mance, FDI, and firm entry as potential pathways through which outages affect job cre- 9 Investments in brownfield (existing) projects could also be affected by outages particularly in the event of prolonged power crisis in a country 10 Under the assumption of free mobility of labor and capital, firms may choose to relocate to areas with reliable supply. However, the cost of relocation is non-trivial. 9 ation in Africa. 3 Data 3.1 Individual and Household Data 3.1.1 Afrobarometer Survey The Afrobarometer survey is a nationally representative survey of public attitudes on democracy, governance, economic conditions, and access to basic social amenities in over 35 African countries. The survey uses a two-stage stratified sampling strategy and fo- cuses on individuals above the age of 18. Data from rounds 6 (2014-15) and 7 (2016-18) of the Afrobarometer are used for the analysis. The dataset is geo-referenced at the com- munity level, making it possible to spatially match it with other datasets. Data from 25 Sub-Saharan African (SSA) countries are used in this paper11 . I use data on employment status of individual(s), quality of electricity supply, so- cioeconomic attributes of the individual and their respective households, and commu- nity characteristics as well. Employment status is measured based on responses to the question: “Do you have a job that pays a cash income?”. Thus employment is defined as equal to 1 if a respondent reports having a cash-paying job, and 0 for a respondent with- out a cash-paying job but actively looking for a job.12 In addition, using the information on their occupational history, respondents were classified into skilled vs unskilled work- ers, and agric vs non-agric sector employees. Quality of electricity supply is measured from the responses of households with electricity connection to the question “how often is the electricity actually available?”. Here the quality of electricity supply is classified 11 Including Botswana, Burkina Faso, Cameroon, Coteˆ d’Ivoire, Cabo Verde, Gabon, Gambia, Ghana, Guinea, Lesotho, Liberia, Madagascar, Malawi, Mauritius, Mali, Mozambique, Namibia, Niger, Nigeria, Senegal, Sierra Leone, South Africa, Swaziland, Zambia, and Zimbabwe 12 In essence, the employment variable is returned missing for respondents that responded “No (not looking)” to the question. 10 as reliable if a (connected) household receives electricity supply always and unreliable if otherwise. Using these data, I compute a measure of reliability in a community based on the share of electrified households in the primary sampling unit that have reliable access to electricity. Specifically, two main measures of reliability are computed: first, a dummy variable (outages in community) equal to 1 if more than 50 percent of connected house- holds in the primary sampling unit (PSU) do not have access to reliable electricity; and second, the share of connected households without reliable access to electricity (outages in community % HH). 3.1.2 Living Standards and Measurement Surveys Household survey data from the Ghana Living Standards Survey (GLSS) and Nigeria General Household Survey (GHS) are used to supplement the analysis in the country- case studies. The GLSS is a nationally representative repeated cross-sectional data. Five rounds of the GLSS data between 1998 and 2017 are used. The GHS on the other hand is a nationally representative household panel data from Nigeria. The analysis relies on three waves of the GHS panel data surveyed between 2011 and 2016. Employment status of respondents is measured in slightly different ways in the two surveys. In the case of the GLSS, individuals are asked about their employment activities over the last 12 months, while in the case of GHS, individuals are asked about their employment activities within the past 7 days preceding the survey. 3.2 Firm Data 3.2.1 World Bank Enterprise Surveys The World Bank Enterprise Surveys (WBES) dataset is a global firm survey that under- takes face-to-face interviews with top managers and business owners in about 145 coun- tries. The survey collects data on several issues relating to firm attributes, access to in- 11 frastructure, constraints to doing business, competition, among others. The survey uses the two-stage stratified random sampling strategy. I use the global standardized version of the dataset which uses a standardized sampling strategy and questionnaire. All mone- tary data are converted into 2009 USD prices using the GDP deflator and exchange rates for the respective countries. The final dataset, therefore, is repeated cross-section data of firms in 10 SSA countries13 surveyed between 2006 and 2018. To account for these time and country variations in the dataset, year and country fixed effects are applied respectively in the estimation. The dataset reports annual revenue and cost of inputs rather than physical measures of outputs and inputs. Productivity in this paper is measured using two indicators: value added per worker and sales per worker. Value added is computed as total sales revenue less the cost of raw materials and intermediate inputs (Hjort and Poulsen, 2019). Addi- tional firm outcomes used paper include number of employees (full-time, temporary), total labor cost, and labor cost per worker. The age of the firm and foreign/domestic own- ership status of firms are also included in the data. Two measures of self-reported power outage intensity are explored in the firm analysis: (i) outage frequency measured as the average number of times a firm experienced power outages in a typical month; and (ii) the number of hours without electricity in a typical month, measured by the product of the frequency and average duration of outages in a typical month. Arguably, these self-reported measures of outage intensity, are not with- out biases. However, administrative data on outage intensity are virtually non-existent in many African countries, thus making the self-reported measures the best possible means of measuring outage intensity. Additionally, given the prevalence and regular nature of power cuts in the study area, the extent of bias associated with recall, if any, will be mini- mal, other things being equal. Finally, the GPS location of firms in the WBES dataset is not publicly available due 13 ˆ d’Ivoire, Congo DR, Ghana, Malawi, Mauritius, Mozambique, Sierra Leone, Togo, Zambia Benin, Cote 12 to privacy concerns. To overcome this challenge, I geo-reference the city/towns in which firms in the dataset are located and match them with other spatial datasets. 3.2.2 Ethiopian Large and Medium Manufacturing and Electricity Industries Survey The large and medium manufacturing and electricity industries survey (LMMIS) is an an- nual census of all large and medium manufacturing firms with at least ten employees and rely on electricity for production in Ethiopia. The data is collected by the Central Statistical Authority (CSA), and all firms that meet the criteria are mandated by law to comply with the requirements of the CSA and participate in the survey (Essers et al., 2021). As a result, the LMMIS captures the universe of all formal large and medium scale manufacturing en- tities in the country (Essers et al., 2021). The main limitation of the data, however, is that it mainly captures formal firms thus excluding firms in the informal sector. The data provides information on inputs, outputs, investments and capital expendi- ture, capacity utilization, duration of operation, as well as the main issues confronting firms. Unlike the WBES, the LMMIS data does not have explicit questions on the intensity of outages faced by firms. Instead, the survey asks firms to list the major issues (including electricity outages) confronting them and in some instances whether these issues affected their ability to operate fully during the calender year. I leverage these responses and mea- sure firms’ exposure to unreliable electricity supply based on whether firms cite outages as a major issue confronting their operation. Specifically, I construct a dummy variable equal to 1 if a firm indicates electricity outages as a major issue either: (i) currently fac- ing the firm; (ii) responsible for not operating at full capacity; or (iii) responsible for not operating all year round; and 0 if otherwise. Data from 2011 to 2017 are used in this paper. Further, using this census data, I compute for each district (Woreda14 ), the number of firms operating in a given year, and firm density (number of firms per 1000 people15 ) for 14 Third administrative region in Ethiopia 15 Subnational population data on Ethiopia were obtained from https://data.humdata.org/dataset/ ethiopia-population-data- -admin-level-0-3 13 each year. These measures provide insights into the distribution of manufacturing firms across space and time in the country. Information on the number of months in the year for which the firm operated is also used in the analysis. 3.3 FDI Data To further understand the effects of outages on entry of (foreign) firms, I use a unique dataset on greenfield foreign direct investment projects. These data are obtained from fDi Markets,16 a subsidiary of the Financial Times (FT) Group. fDi Markets database tracks cross-border investment projects around the world since 2003. The database is primarily used by agencies such as the World Bank, the Economist Intelligence Unit and UNCTAD in monitoring cross-border investments. It provides granular data on the project’s primary sector, sub-sector, country (city) of origin, destination country (city), investment size, etc. For the purpose of this paper, FDI project data on 23 emerging markets in Africa, Latin American and Caribbean, and Asia17 between 2007 and 2017 focusing solely on investment in sectors excluding energy and construction sectors. 3.4 Lightning Data Lightning intensity is used as an instrument for electricity outages. However granular data lightning occurrence in Africa and many developing countries is a challenge. Available data on lightning activities in developing countries are mainly satellite-based measures of lightning intensity which come from NASA’s LIS/OTD Gridded Lightning Climatology Dataset.18 . This dataset is cross-sectional and reports the average number of lightning strikes between 1995 and 2010 and is available at a relatively low spatial resolution of 16 https://www.fdimarkets.com/ 17 Cambodia, Ghana, Guatemala, Honduras, India, Jamaica, Kenya, Morocco, Mexico, Myanmar, Mau- ritius, Namibia, Nicaragua, Pakistan, Philippines, Senegal, South Africa, Uganda, Uruguay, Uzbekistan, Vietnam, Zambia, Zimbabwe 18 https://ghrc.nsstc.nasa.gov/uso/ds docs/lis climatology/LISOTD climatology dataset.html This is arguably the most widely used measure of lightning intensity in the literature 14 0.5◦ × 0.5◦ . The lack of temporal variation due to the cross-sectional nature of the dataset coupled with relative low spatial resolution (0.5◦ × 0.5◦ ) poses a challenge to the use of this dataset as an instrument to causally estimate the effects of outages. Recent scientific literature however shows that observed lightning intensity is propor- tional to the product of convective available potential energy19 (CAPE) and precipitation rate (i.e., the amount of precipitation that would cover a given area (m2 ) per second) (Romps et al., 2014; Dewan et al., 2018). In an attempt to find granular data on lightning intensity and aid future projections of lightning intensity associated with climate change, Romps et al. (2014) proposed a proxy for measuring lightning intensity: η F = × CAP E × P (1) E where F is the lightning flash rate per area (m−2 s−1 ), while CAPE (J kg−1 ) and P (kg m−2 s−1 ) represent the convective available potential energy and precipitation rate respec- tively. η/E , a constant of proportionality, is the ratio of the conversion efficiency factor (η ) and the energy discharge (in joules) per flash (E ).20 Romps et al. (2014) validate this methodology by showing that CAP E × P explains about 77% of the variance in actual lightning flash rate in continental United States. Dewan et al. (2018) also using data from Bangladesh show a significant correlation between actual lightning strikes and CAP E × P with the latter explaining about 89% of the variance in the former on a monthly time scale. Therefore, following Romps et al. (2014) and Dewan et al. (2018), I use CAP E × P as a proxy for lightning intensity. Specifically, using time series data on CAPE and precipi- tation rate from the ERA5 Global Reanalysis Database by the Copernicus Climate Change Service21 , I compute CAP E × P at a 0.25◦ × 0.25◦ grid-cell level and use it as an instrument 19 This measures the amount of energy a parcel of air would gain if raised to a specific height in the atmosphere. See: https://study.com/academy/lesson/ convective-available-potential-energy-cape-definition-use-in-forecasting.html 20 See Romps et al. (2014) for details. 21 https://cds.climate.copernicus.eu/cdsapp#!/home 15 for the level of electricity outages in the respective years. To demonstrate that this proxy correlates with actual lightning intensity in the African context, I use data on lightning flash rate from NASA’s OTD/LIS satellite and correlate it with the proxy (CAP E × P ) as shown in Figure A1, and Figure A2 in the online appendix. The scatter plot in Figure A1 reveals a high r-square: CAP E × P explains about 77% of the variations in lightning intensity in Africa, thus confirming the earlier findings of Romps et al. (2014). Figure A2 in the online appendix also shows a high spatial correlation between the actual lightning flash rate and the proxy. In addition, data on mean annual temperature and precipitation from the ERA5 Global Reanalysis Database22 are used in the analysis. Summary statistics are presented in Table A1 in the appendix. 4 Empirical Strategy and Results 4.1 Identification Strategy Empirical estimation of the causal impact of electricity outages on outcomes such as em- ployment and firm productivity is often beset with methodological challenges. Notable among them is the issue of endogeneity resulting from the potential correlation between outage intensity and (observable and unobservable) factors that (in)directly influence these outcomes. In other words, any assumption that variations in outage intensity are orthogonal to economic outcomes such as employment and firm productivity is unlikely to be valid. For instance, firm location, industry composition, and the prevailing economic and political conditions in a country can influence both outage intensity and the perfor- mance of firms. Also, regions with high unemployment and hence low income are more likely to suffer outages possibly due to the non-payment of electricity services leading 22 https://cds.climate.copernicus.eu/cdsapp#!/home 16 to the vicious cycle of outages and non-payment of electricity bills (Dzansi et al., 2018). Additionally, self-reported measures of outage intensity are plausibly measured with er- ror, hence the possibility of a downward bias (attenuation bias) in the impact from OLS estimation cannot be ignored (Allcott et al., 2016). To address these issues, I utilize the in- strumental variable approach and exploit spatial and time variations in lightning intensity as an instrument for power outages. Lightning is a major cause of power outages around the world particularly within the tropics where thunderstorm activities are prevalent (Andersen et al., 2011, 2012; Andersen and Dalgaard, 2013). Lightning strikes contain about a billion volts of electricity; there- fore when it strikes a transmission line or transformer, it induces voltage surge, thereby destroying the transmission lines and equipments23 and curtailing the flow of electricity. Electrical infrastructure destroyed by lightning induced voltage spikes and dips, could take several days to be repaired and often entail high cost of replacement. As a result, affected communities often go several days without electricity. In South Africa for in- stance, lightning is estimated to account for nearly 65% of all over-voltage damages to electrical transmission network24 , with strikes within 40 meters of a transmission (distri- bution) line causing significant damages (Andersen and Dalgaard, 2013). Similar effects have been recorded in Swaziland25 (Mswane and Gaunt, 2005), Nigeria26 and Ghana27 . In the United States, lightning activities account for about a third of all incidence of power outages (Chisholm and Cummins, 2006). Moreover, the fact that lightning activities are natural phenomena, it induces random variations in the incidence and intensity of outages across space and time conditional on locational and climatic characteristics. Consequently, lightning strikes have increasingly been used as an instrument for the quality of electric- 23 http://www.liveline.co.za/lightning-stats.php 24 http://www.liveline.co.za/lightning-stats.php 25 Lightning accounts for about 50% of outages incidence in the country (Mswane and Gaunt, 2005) 26 Adepitan and Oladiran (2012) estimates that lightning accounts for nearly 10% of random electricity outages experienced in Ijebu province. 27 http://www.ghanaweb.com/GhanaHomePage/NewsArchive/Lighting-cuts-electricity-to-Sissala- East-district-453319 17 ity and diffusion of electric-powered technologies in the economics literature (Andersen et al., 2011, 2012; Andersen and Dalgaard, 2013; Manacorda and Tesei, 2020; Guriev et al., 2020). Therefore, I estimate the effects of electricity outages on employment using the IV framework with the baseline regression specified as follows: IV first-stage: ′ Outagesjct = ϕ × Lightningjct + Xijct α1 + θc + δt + µijct (2) IV second-stage: ′ Yijct = β × Outagesjct + Xijct α2 + θc + δt + ϵijct (3) where Yijct is the outcome variable for individual i living in community j , country c at time t. Outagesjct is a measure of (average) electricity outage intensity in the community. Two measures of outage intensity are explored here: first, an indicator variable equal 1 if more than 50% of households (respondents) interviewed in the primary sampling unit (PSU) receive poor quality electricity services28 in the relevant period and 0 if otherwise; Second, the share of households who experience poor (quality) electricity services. The communal measure of outage intensity is preferred to household measure as the former captures, to a large extent, the general quality of electricity services in the community relative to the latter. More so, since most individuals are usually employed outside their home29 , a household measure of outage intensity may not suffice. Additional estimations are done using measures of outage intensity at the district level to account for the effects of outages on employment of people who work outside their communities. The results 28 Poor quality electricity is defined as either having electricity occasionally, about half of the time or most of the time. Electricity supply is defined as of high quality if a connected household receives electricity all the time. In Section 4.2.3, I explore alternative estimations by redefining our measure of electricity reliability to include households who receive electricity most of the time and always. 29 Even for individuals employed in household enterprises, the enterprises may be located outside the home (e.g. in market or trading centers). Hence the average quality of electricity supply in the community suffices as a good measure of than the quality of supply reported at the household level. 18 remain robust to the measure of outages. Xijct is a vector of individual controls including age and gender. I also control for temperature and precipitation to absorb the potential channels through which the instrument (lightning) could affect the outcome variable. θc and δt represent respectively, country and year fixed effects to control for time-invariant characteristics as well as time-varying correlates of the outcome variable. Lightningjct represents the lightning intensity in community j at time t. As highlighted in Section 3.4, this is proxied by CAP Ejct × Pjct . The exclusion restriction assumption requires other than through outages, unemploy- ment rates should not vary over time across locations depending on the average lightning intensity. In other words, lightning intensity is not correlated with the outcome variable via channels other than through outages. While this assumption is not directly testable, I identify two possible channels through which this assumption may be violated. First, given that lightning strikes are associated with rainfall, and rainfall also being a driver of economic activities particularly in developing countries where agriculture is the main stay, lightning could influence employment outcomes via rainfall. To mitigate this con- cern, I control for local climate conditions such as temperature and precipitation (rainfall) directly in the regression to absorb this potential channel. Secondly, lightning can directly (indirectly) also influence the diffusion of modern technologies such as mobile phones. As shown by Manacorda and Tesei (2020) and Guriev et al. (2020), areas with high light- ning intensities often tend to have low penetration of digital technologies such as mobile phones as voltage surges often associated with lightning causes damages to electrical com- ponents of digital infrastructure. Thus to the extent that access to digital infrastructure (technology in general) matter for productivity and employment outcomes, lightning ac- tivities may possibly influence employment through uptake of these technologies. Again, to mitigate this concern, I control for the extent of technology diffusion in the respective communities (cities) using mobile phone (2G/3G/4G30 ) coverage rate as a proxy. Thus 30 second, third and fourth generation mobile technologies 19 conditional on the technology diffusion, climate indicators, and the other controls includ- ing location and time fixed effects, I argue that lightning intensity influence employment outcomes only through its effect on electricity outages. µijct and ϵijct represent respectively, the error terms for the first and second-stage equations. Standard errors are clustered at community level (primary sampling unit). The main parameter of interest, β measures the causal effect of electricity outages on ˆIV recovers the the outcome variable. Therefore conditional on the instrument validity, β local average treatment effect (LATE) of electricity outages on the outcome variable(s). 4.2 Baseline Results This section presents the baseline results on the effect of outages on employment using the Afrobarometer dataset from several African countries. 4.2.1 First-Stage IV Regression Table 1 presents the results of the first-stage regressions which estimate the relationship between the lightning intensity and our measures of outages: whether a community ex- periences outages (columns 1-2) and the share of households in a community (columns 3-4) experiencing outages. The results show a positive association between lightning intensity and outages. In col- umn 2 for instance, I find that a percent increase in lightning intensity is associated with a 12 percentage point (pp) increase in the probability of a community experiencing outages. Similarly in column 4, a percent increase in lightning intensity is associated with a 9 pp increase in the share of households experiencing outages. The strength of the instrument is relatively high as the first-stage F-statistic (Fstat) exceeds the conventional benchmark of 10 (Stock and Yogo, 2005) in all specifications. To complement the results in Table 1, Figure 1 presents a binscatter plot relationship be- 20 tween lightning intensity and the two measures of outages. The plot confirms the (strong) positive association between outages and lightning activities. 4.2.2 Main IV Results Table 2, reports the OLS, second-stage IV and reduced formed estimates. I estimate two variant specifications by alternating between survey year and round31 fixed effects. Start- ing with OLS estimates column 1, the results show a negative association between expo- sure to electricity outages and employment outcomes. For instance, in column 2, living in a community with outages is associated with an increase in the probability of being unemployed by 2 pp. Likewise, the effect is negative and statistically significant in rela- tion to non-agric employment and unskilled jobs. The relationship between outages and employment of skilled and agric-sector workers are statistically insignificant. While these estimates are non-causal they provide suggestive evidence that outages are negatively as- sociated with lower employment outcomes. Turning to the IV estimates, the results also show that outages indeed have negative effects on employment, albeit the IV estimates are relatively large compared to the OLS es- timates.32 In column 2 for instance, I find that living in a community that experiences fre- quent electricity outages reduces the probability of employment by 13.5 pp. Interestingly, it appears that these effects are driven entirely by the effect on employment in non-agric sectors (column 3-4) as the effect on the probability of employment in agric related jobs is almost zero (5-6) and statistically insignificant. There are at least two candidate rea- sons behind the null effect of outages on agric-related employment: first, the agric sector in many African countries have low technology intensity, as a result, the potential effect of electricity outages on production in the sector is minimal. Secondly, given the defini- tion of employment as measured in the Afrobarometer as being having a “cash paid” job, 31 The differences arise because survey rounds often overlap calender years across countries 32 This suggest the possibility of a downward bias in the OLS estimates plausibly due measurement errors in the measure of outages Allcott et al. (2016). 21 only 6.1% of the individuals in the sample with such jobs are employed in the agric sector. In other words the sample of employed individuals in the sample are heavily skewed (≈ 94%) towards the non-agric sectors which incidentally are more energy intensive than the agric sector. Hence, it is unsurprising to see a null effect of outages on employment in the agriculture sector. How does the employment effect of outages vary between skilled and unskilled work- ers? To address this question, I split the sample into skilled and unskilled workers and es- timate the baseline equations. In columns 7-8, the results show a negative and statistically significant negative effect of outages on employment of skilled workers. Outages reduce the probability of employment of skilled workers by 19 pp (column 8). Interestingly, the effect of outages on employment of unskilled workers are statistically insignificant albeit negative. The null effect of outages on unskilled jobs provides a good placebo test to the validity of the IV design, showing that the IV regressions are only picking up variations in lightning intensity that affect electricity outages as we do not expect outages to signifi- cantly alter the labor market outcomes of people in low skilled occupations given the low energy intensity of such occupations. Thus, if the instrument is picking-up other effects such as the diffusion of digital technologies like mobile phones, one would expect to see robust negative effects on employment outcomes of both skilled and unskilled persons as well as agriculture-sector jobs as mobile phones have been shown to be positively cor- related with household welfare in developing countries (Bahia et al., 2020; Masaki et al., 2020). Still, in Table 2, I present the reduced-form estimates showing the relationship be- tween lightning intensity and the probability of employment. Unlike the IV estimates, the reduced-form estimates do not require the exclusion restriction assumption to hold. The results show a negative association between lightning intensity and employment out- comes: an increase in lightning intensity is associated with a lower probability of employ- ment, particularly, in the non-agric sectors and skilled jobs. 22 Taken together, the results from the table suggest that while outages constrain em- ployment, skilled workers are disproportionately affected. This is plausibly due to the fact electricity is a key input in the production process, and thus essential for most skilled workers in undertaking their activities at the workplace. Hence, unreliable electricity pro- vision has negative consequences on the creation of skilled jobs in Africa. 4.2.3 Sensitivity analysis and additional results In this section, I present additional analyses of the effects of outages across heterogeneous groups and also explore the robustness of the baseline results to measure of exposure to outages. Private vs Public Sector Jobs: In addition to the effects of outages on employment across skill levels, and agric vs non-agric sector jobs, I also explore the effects across employ- ment in private vs public sectors (see Table A2 in the online appendix). This distinction is important for at least two reasons. First, as profit maximizers, private firms are likely to lay off workers when faced with challenges such as an energy crisis which increases their production costs and lowers profits. However, public sector firms may be able to keep workers on their payroll even in times of crisis, for instance, due to their ability to receive government subsidies for job preservation. Secondly, in many developing countries, the public sector remains the largest employer of formal sector workers33 and job cuts in this sector come with political costs.34 In some cases, public sector jobs are likely to increase during periods of crisis as a way to reduce unemployment.35 Table A2 (online appendix) presents the results on the effects of outages on the prob- ability of employment: all sectors (columns 1-2), private sector (columns 3-8), and public sector (9-10). In column 4, the IV estimates suggest that outages have a substantial nega- 33 https://blogs.worldbank.org/arabvoices/governance-and-public-sector-employment-middle-east-and-north-africa 34 https://www.imf.org/external/pubs/ft/fandd/1998/06/lienert.htm 35 https://www.imf.org/external/pubs/ft/fandd/1998/06/lienert.htm 23 tive effect (≈ 30 pp) on the probability of employment in the private sector (i.e., working for a private firm or self-employment). However, this effect largely pertains to employ- ment by private firms, as there is no effect on self-employment. Interestingly, contrary to the negative effect on employment by private firms, the effect on being employed in the public sector (column 10) is positive and sizeable (≈ 15 pp). This is plausibly sug- gestive of the role of state-owned enterprises in job preservation in developing countries. Juxtaposing the results in columns 3-8 with 9-10, one can conclude the baseline effects in columns 1-2 are likely the overall net effects of outages on employment. Effects Across Gender: In Table A3, I also explore the employment effects of outages across gender, and do not find any consistent evidence of significant gender differences in the impact. In other words, outages reduce the employment outcomes of workers irre- spective of their gender. Measurement of outages: In the baseline analysis, outage is defined as an indicator vari- able equal to 1 if at least 50% of respondents with electricity connection in the community (PSU) report having unreliable supply of electricity and 0 if otherwise. To test the sensi- tivity of the analysis to this measure, I explore two approaches. First I define variant measures of the outage dummy by alternating the threshold of the (minimum) share of respondents with unreliable supply of electricity. Specifically, I define additional outage dummy variables coded 1 if at least 10%, 20%, 30%, 40%,..., 90% of respondents with electricity connection report having unreliable supply of electricity and 0 if otherwise. Using these dummies, I estimate separate regressions using the base- line specifications to assess the sensitivity of the point estimates to these measures. The results in Figure A3 in the online appendix show robust negative effects of outages on employment rates. The respective estimates for the dummy variables between the 10% and 70% minimum thresholds are relatively stable, negative and statistically significant at 24 95% confidence interval. The corresponding estimates for the outage dummies based on the 80% and 90% minimum thresholds are however relatively large and significant only at 90% confidence interval. As a second strategy, I estimate a variant model using the share of respondents experi- encing unreliable electricity supply instead of using the discrete measurement of outages. Results are shown in Table A4 in the online appendix. Once again, the results are quali- tatively and quantitatively similar to the baseline results that outages have economically and statistically significant negative effects on the probability of employment. Exposure to Outages at the District Level: A possible critique of the above analysis is that outages at the community level may not capture the full impact of exposure to unre- liable electricity supply, particularly for individuals who work outside their communities. To address this concern, I compute a variant measure of exposure to outages using the share of households in a district (second-level administrative region) experiencing with an unreliable supply of electricity. Essentially, I compute a dummy variable “Electricity Outages in District (0/1)” defined as equal to 1 if more than half of respondents in the district experience outages and 0 if otherwise. Using this measure, I replicate the baseline estimation in equations (2) and (3) to assess the effects of outages at the district level on the probability of employment.36 The corresponding first and second-stage IV results are shown in Tables A5 and A6 in the appendix respectively. Once again, the IV estimates in Tables A6 are qualitatively and quantitatively similar to the baseline results in Tables 2. This provides an additional assurance that the results are not driven by measurement errors in the outage measure. Role of Electrification Rates in Measurement of Outage Intensity: Given that our mea- 36 Accordingly the instrument used here is the log of the average lightning intensity in the district. Stan- dard errors are clustered at the district level. 25 sure of outages at the community level is conditional on the share of households with electricity connectivity, a potential concern is that the estimated effect of outages on em- ployment could be picking-up the effects of electrification, rather than the ”pure effect” of outages. Admittedly, even if this is the case, this could potentially imply that our estimates are lower bound, given the evidence in prior literature showing that access to electricity is associated with increased employment (Dinkelman, 2011). Nonetheless, I leverage two strategies to show that the baseline results are not driven by differences in electrification rates. First, in Table A7, I explicitly control for electricity access by including a dummy variable set equal to 1 if the electricity access rate in the community is above the median and 0 if otherwise.37 The results remain robust, showing that indeed exposure to outages is associated with a decline in employment. Second, I restrict the sample to communities with universal access to electricity. Intuitively, by so doing, I net out the effects arising from differences in electrification rates across communities. Once again, the results in Ta- ble A8 confirm that outages are associated with a decline in employment. Role of Outliers: An additional concern relates to the sensitivity of the first-stage relation- ship to the inclusion or exclusion of countries. In other words, is there a particular set of countries for which the lightning-outage relationship holds? For instance, is it that coun- tries with low grid quality are the ones where outages are sensitive to lightning strikes? And if yes, is unemployment peculiar to such countries? The sensitivity of our first-stage results to the omission of countries could pose questions to the IV results. To this end, I perform a ”leave-one-out” exercise by estimating the first-stage equation while system- atically excluding each of the sample countries one at a time. The results in Figure A4 suggest that our results are not driven by the inclusion/omission of specific countries, as the results remain largely stable to the sample composition. 37 It is important to emphasize that given the relationship between electricity connection and outages, controlling for the degree of electrification could lead to the problem of ”bad controls” (Angrist and Pischke, 2009) 26 4.3 Country Case Studies In this section, I undertake a deep-dive analysis of the effects of outages on employment by presenting two country case studies: Ghana and Nigeria 4.3.1 Ghana: The ”Dumsor” Power Crisis (2012-2016) Between 2012(Q4) and 2016, Ghana faced with the worst power crisis in the country’s his- tory. The crisis was so severe that it led to massive demonstrations by citizens protesting38 against the government’s inability to provide reliable power. The incessant power cuts during the period was nick-named ”Dumsor39 ” to wit ”off and on”. To manage the crisis, the Electricity Company of Ghana40 (ECG) implemented a power rationing program where available power were rationed among communities using a sched- ule published in the newspapers (see Figure 2). At the height of the crisis, consumers were guaranteed only 12-13 hours of electricity for every 36 hour period. The effect of the crisis on firms and industry was severe leading to significant produc- tivity losses, particularly, among small-and-medium scale enterprises, particularly those without access to in-house generators as a backup option (Abeberese et al., 2021). To what extent did this crisis affect employment? The power crisis in Ghana provides a unique quasi-natural experiment to examine the effect of electricity outages on employment. To this end, I leverage household survey data between 1998 and 2017 and exploit plausi- bly exogenous variations in exposure to power crises to estimate the effect of outages on employment using a difference-in-difference (DiD) design (Kuka et al., 2020; Verner and Gyongy ¨ ¨ 2020): Essentially, I exploit differences in dependence of local economies on osi, electricity and estimate the differences in employment outcomes between high and low electricity dependent districts before and after the crisis using the specification in equa- 38 https://www.theguardian.com/world/2015/may/17/ghanas-celebrities-lead-protest-marches-against-ongoing-energy Accessed: December 2020 39 https://en.wikipedia.org/wiki/Dumsor 40 the main distributor 27 tion 4 ′ Yidt = ϕ × HighExposured × P owerCrisist + Xidt α + γWd × t + θd + δt + λY OB + ϵijct (4) where Yidt is the employment status of individual i, in district d , surveyed in year t. HighExposure is a measure of a district’s dependence on electricity before the crisis. It is defined as an indicator variable equal to 1 if electricity access rate in the district at the base- line41 is higher than the median access rate in the country and 0 if otherwise. P owerCrisis is an indicator variable equal to 1 for the period between 2013 and 2016 when the power crisis was at its peak. Xidt is a vector of individual controls. I also control for trends in observable determinants of local economic development by including district level charac- teristics (during the baseline) interacted with linear time trend. District and year fixed ef- fects are represented by θd and δt respectively. I also include birth-year fixed effects (λY OB ) to account for age-cohort effects. Standard errors are clustered at district level. The coefficient of the interaction between HighExposure and P owerCrisis represented by, ϕ, measures the difference in employment outcomes between high and low access (electricity dependent) districts in the crisis and non-crisis periods. The intuition behind the identification strategy is that economic activities in districts with high electricity ac- cess are plausibly highly reliant (dependent) on electricity for economic activities relative to low access districts. As a result, an exogenous shock to electricity supply is likely to affect high access districts relative to low access districts. The validity of this research design relies on the assumption that absent the electricity crisis, trends in the outcome variable (employment) between low and high access (exposed) districts are parallel. In other words, the change in employment rate between the treated and control districts is uncorrelated with underlying (observable/unobservable) trends prior to the treatment. 41 I use data from the 2000 population and housing census to compute the access rate in each district. 28 To test for pre-trends and the evolution of employment rates over time, I estimate an event study: ′ Yidt = ϕτ × HighExposured × I(year = t) + Xidt α + θd + δt + λY OB + ϵijct (5) τ ̸=t−ι where I(year = t) is an indicator variable that equals 1 in year t and 0 if otherwise. All else remains as previously defined. Results In evaluating the effect of the power crisis on employment, I leverage two survey datasets - GLSS and Afrobarometer - spanning over the period 1998 and 2018. Table 3 presents the DiD estimates of the effect of the power crisis on employment. For each dataset, I estimate three variant specifications to test for the robustness of the esti- mates to various controls. My preferred specification is column 3 (6) which is the most restrictive: including fixed effects for district, survey year, and birth-year. Individual and community characteristics such as gender, educational attainment, rural/urban status, ac- cess to road and water are also included. In addition, I control for the interaction between linear time trends and district-level characteristics such as the average homeownership and literacy rates at the baseline. These allow us to absorb trends in local economic devel- opment that may be correlated with unemployment rates across districts. The results are stable and consistent across the two datasets. Starting with the GLSS dataset, the results in column 3 suggest that the crisis led to a 3.4 pp reduction in the probability of employment compared to a 4.7 pp reduction based on the Afrobarometer dataset (column 6). These are economically and statistically significant. Next, I explore the trends in (un)employment rate before, during and after the cri- sis. These dynamics are not only important for examining parallel trends assumption, but 29 also, the response in employment after the crisis. Results from the event study based on data from the GLSS are shown in Figure 3. First, I do not find any statistically significant differences in the trends in employment between high (treated) and low (control) access districts prior to the crisis. This provides support to the identification strategy that the es- timated drop in employment during the crisis is uncorrelated with underlying differences between treated and control districts prior to the crisis. Secondly, between 2013 and 2016 (the crisis period) employment rates fell on average42 by 7.2 pp relative to the reference period (2006).43 Interestingly, the employment rate in 2017 (a year after the crisis ended), was about 5.6 pp lower than the reference year: an indication of a slow recovery rate in employment– and plausibly the economy– after years of exposure to the crisis. The foregoing analysis provides additional causal evidence of the negative effects of electricity outages on employment. 4.3.2 Nigeria Nigeria is among the leading African countries with the lowest levels of reliability in elec- tricity supply. According to data from the World Bank Enterprise Survey, about 77.6% of Nigerian firms report experiencing outages44 compared to the SSA average of 75.6. The number of power outages experienced by Nigerian firms in a typical month is 32.8 com- pared to the SSA average of 8.5. Given this, to what extent is unreliable electricity supply constraining employment in Nigeria? I leverage unique household panel data from the Nigerian General Household Sur- vey (GHS) to explore the effect of unreliable electricity provision on employment in the country using a fixed effect model specified as follows: 42 the estimates for 2013 and 2016 are 6.7 pp and 7.2 pp respectively 43 Since the crisis started briefly in the fourth quarter of 2012, I decided against using 2012 as the reference period. 44 Based on 2014 data. See https://www.enterprisesurveys.org/en/data/exploreeconomies/2014/ nigeria#infrastructure 30 ′ Yihjst = ϕ × Outagesj (h)st + Xihjst α + θh + δs×t + ϵihjst (6) where Outagesj (h)st is coded 1 if a household indicates frequent electricity outages45 in the community and 0 if otherwise. In other specifications, I also explore alternate measures using the reported frequency of outages in the community. θh and δs×t represent house- hold and state ×year fixed effects respectively. In this setup, identification relies on within household variations in exposure to electricity outages. Thus while the distribution of out- ages may plausibly be non-random, within household variations in exposure to outages in their communities could be plausibly random. This assumption is arguably strong and thus the estimated effects may not be strictly causal; they nonetheless offer insights on the effects of outages conditional on household and state ×year fixed effects. Standard errors are clustered at the level of primary sampling unit. Results Table 4 presents the fixed effect estimates of the relationship between electricity outages and employment in Nigeria. In columns 1 and 4, I estimate the baseline specification with- out any household controls. Columns 2 and 5 include the full set of controls such as gen- der and educational attainment, as well as the rural/urban status of the community. In columns 3 and 6, I exclude the northeastern part of Nigeria from the analysis. Terror- ist group Boko Haram46 has since 2002 staged insurgency attacks in north-eastern Nigeria. This has created insecurity in the region with negative socioeconomic impact (Bertoni et al., 2019). Therefore to isolate the effects of the terrorist activities from contaminating the results, I exclude households in these areas from the estimations in columns 3 and 6. Columns 1-3 present the results on the relationship between electricity outages and the probability of employment using a dichotomous measure of outages: defined 1 if a 45 either daily, several times a week, several times a month or several times a year 46 https://en.wikipedia.org/wiki/Boko Haram 31 household reports experiencing outages in the community either daily, or several times a week/month/year, and 0 if otherwise. The results in column 1 (2) indicate that outages in the community are associated with a 5.7 (5.9) pp reduction in the probability of employ- ment, and are statistically significant at 5% (10%) error level. The effect remains negative and statistically significant (5% error level), albeit slightly higher (in absolute terms) in column 3 when I exclude households from northeastern Nigeria. In columns 4-6, I explore how the effects vary according to the intensity (frequency) of outages. The results show that outages have negative and statistically significant effects on employment in communities that experience outages either daily, several times a week, or several times a month, relative to the reference category (those who do not experience outages in their community). The effects for those who experience outages several times a year in their communities are however not statistically significant relative to the reference category. Although causal interpretations of the fixed effects estimate require strong assump- tions of exogeneity, the estimates are qualitatively and quantitatively to the DiD estimates from Ghana. Thus the results herein, once again, provide additional suggestive evidence that outages are associated with high rates of unemployment. 5 Mechanisms This section presents evidence on the channels through which electricity outages affect employment. I explore these channels along the extensive and intensive margins. On the extensive margin, distortions in the business environment like electricity outages have the potential to discourage potential entrepreneurs to establish new enterprises due to the perceived constraints to doing business. In the intensive margin, electricity outages reduce firm performance. As a result, existing firms may either reduce their labor demand or reduce wages with potential implications on employment. 32 5.1 Intensive Margin Here, I examine the effects of outages on firm performance and consequently firms’ de- mand for labor and labor cost as potential channels underlying the outage-employment nexus. Using firm-level data from the WBES, I estimate the following IV specification: IV first-stage: ′ Outagesikct = ϕ × Lightningkct + Xijct γ1 + θc + δd×t + ϵikct (7) IV second-stage: ′ Yikct = β × Outagesikct + Xikct γ2 + θc + δd×t + µikct (8) where Yikct is the outcome of firm i in city k , country c, and year t. Outagesikct is the outage intensity experienced by firm i. Two measures of outage intensity are explored here: the number (frequency) of outages a firm experiences in a typical month, and the total number of outage hours experienced by the firm in a typical month. θc and δd×t ′ represent country and industry×year fixed effects respectively. Xikct is a vector of firm, climate, and city controls including mobile phone coverage rate as a proxy for the rate of technology diffusion at the city level. Lightningkct is a proxy of the average lightning intensity in city k at time t. Again as highlighted in section 3.4, I use CAPE× P as a proxy for lighting intensity. Standard errors are clustered at the city level. 5.1.1 Firm Performance and Labor Demand Table 5 presents the OLS, IV, and reduced-form estimates of the effect of outages on firm performance.47 The results show a negative and statistically significant impact of outages on sales, sales per worker, and value-added per worker. The findings are consistent with 47 Table A11 in the online appendix presents the first-stage results 33 the frequency and duration of outages. For instance, in column 2 (4), a percentage increase in the number (hours) of outages experienced by a firm reduces sales by 1.2% (0.6%). Sim- ilarly, a percent increase in the number (hours) of outages experienced by a firm reduces sales per worker by 1.3% (0.7%) (column 6 (8)). The effects on value added per worker are also negative and statistically significant: a 2.3% (2.5%) reduction in value added per worker for every percent increase in the number (frequency) of outages experienced by the firm. Arguably, these estimates can be regarded as lower bound estimates of the total effect of outages given the fact that some firms in the dataset rely on electricity self-generation during periods of blackouts to mitigate the impact of outages. Disentangling the extent of attenuation in the impact of self-generation is, however, an empirical challenge as the deci- sion to self-generate and the degree of self-generation are plausibly endogenous. Nonethe- less, the evidence herein unambiguously highlights the challenges of African firms48 in operating in a business environment with unreliable access to power. Given the above negative effect of electricity outages on productivity, what are the im- plications on labor demand by African firms? Table 6 presents results on how electricity outages affect the number of workers hired by a firm and the associated labor cost49 . I ex- plore the effects of outages on the number of full-time and temporary workers employed by firms in columns 1-4 and 5-8 respectively. The effects of outages on employment of full- time (permanent) staff are negative but statistically insignificant. The effects on temporary (part-time) workers are however negative and statistically significant across all specifica- tions: a one percent increase in the number (hours) of outages experienced by the firm is associated with a 0.6% (0.3%) reduction in the number of temporary staff employed by the firm (column 6(8)). The differences in the effects of outages on the employment of permanent and temporary staff could be associated with constraints in the labor market 48 The term “African firms” as used in this paper, refers to firms operating in Africa, without any conno- tation to the nationality of its owners or country of origin. 49 See Table A12 in the appendix for the corresponding first-stage results 34 such as unionization and labor laws that offer safeguards to (permanent) workers against indiscriminate layoffs. As a result, employers may resort to measures such as wage rene- gotiation or reduction in the number of working hours to be able to manage the negative impact of unreliable electricity provision on their activities. Substantiating these argu- ments requires granular data on the average working hours by the workers of firms in our sample, as well as data on the workers’ wages. However, data on these measures are un- available in the WBES. Nonetheless, I use data on total labor cost and labor cost per worker as proxies for wage and explore its relationship with outages experienced by firms. Interestingly, the effects on total labor cost and labor cost per worker are negative and statistically significant. A percent increase in the number (hours) of outages experienced by firms is associated with a 1.1% (0.6%) reduction in labor cost (column 10 (12)). Sim- ilarly, the labor cost per worker reduces by 1.15% (column 14) and 0.64% (columns 16) respectively for every percent increase in the number and hours of outages experienced by a firm: perhaps an indication of lowering wages in response to the negative effects of outages on productivity. I also explore how the effects of outages on firm performance vary across firm size,50 and energy intensity. In Table A13 for instance, the results suggest that the negative ef- fects of outages on firm performance pertain largely to small firms. The effects on medium and large firms are not statistically significant albeit negative. These results are perhaps indicative of the ability of large and medium-sized firms to invest in abatement technolo- gies such as the use of generators or reliance on captive power generations to mitigate the unreliable grid electricity supply on their operations. Similarly, the results in Tables A14 and A15 also suggest that the outage impacts on firm performance and labor demand is largely concentrated among firms in high energy-intensive sectors. This is plausibly due to the critical role of electricity in the daily operations of firms in such sectors. 50 Definition of firms are as follows: small firms (¡20 employees), medium (20-99 employees), and large (100 and above employees) 35 Overall, the results in Table 5 provide suggestive evidence that outages negatively af- fect the performance of firms. Firms respond to this negative shock by reducing demand for temporary workers, and reducing wages so as to avoid (massive) lay-offs of perma- nent full-time staff (see Table 6). Available evidence from the relatively scant literature on the effects of electricity outages on firm productivity and labor demand lends support to the findings of this paper (Allcott et al., 2016; Hardy and McCasland, 2019; Abeberese et al., 2021). Allcott et al. (2016) for instance provide evidence of significant revenue and productivity losses resulting from electricity outages in India. Also using data on small garment firms in Ghana, Hardy and McCasland (2019) show that the effect of electricity outages on firm revenue and profitability is non-trivial. Hardy and McCasland (2019) further show that firms respond to these shortages by reducing production hours without any allocation to non-outage days, and more importantly substituting high-wage employ- ees for low-wage employees. 5.2 Extensive Margin In this section, I provide evidence on how unreliable electricity provision constrains job creation on the extensive margin by showing the effects on: (i) (net) entry of firms using data from Ethiopia, and (ii) entry of foreign firms via FDI exploiting the “Dumsor” crisis in Ghana as a natural experiment. 5.2.1 Firm Entry The entry of new firms is an important channel for job creation as it leads to the expansion of the productive sectors and employment opportunities. Examining the effects of unre- liable power provision on entry (and exit) of firms is however a challenge mainly due to the lack of administrative data on firms. In many African countries, firm census are scant, with the exception of Ethiopia and South Africa that conduct yearly censuses of firms. An 36 additional challenge is that even where firm census data are available, there is a dearth of information on the quality of electricity supply, thus constraining empirical assessment of the role of electricity shortages on (net) entry of firms. In this section, I use firm census data from Ethiopia to show the relationship between exposure to outages and the density of (manufacturing) firms across locations in the coun- try and also show how outages influence the operation (shutdown) of (manufacturing) firms in the country. Specifically, I estimate the following fixed effect specification ′ Yit = ϕ × U nreliableSupplyit + Xit α + θi + δt + ϵit (9) where Yit is a placeholder for firm outcomes in location (district or city) i at time i, U nreliableSupplyit ′ is a measure of unreliable electricity provision at location i in time i, Xit is a vector of con- trols, while θi and δt represent location and time fixed effects respectively. The analysis is conducted at district and firm levels. In the district-level analysis, two main outcomes are explored: the number of firms in a district (Woreda) in a given year, and firm density (i.e. the number of firms per 1000 people). These outcomes represent the (net) entry of firms across various districts in Ethiopia in a given period. Thus, using these outcomes, I explore the potential effect of unreliable electricity provision on the intensity of (manufac- turing) firms across locations. A priori, unreliable electricity provision in a given locality is expected to influence the exit (and entry) of firms thereby resulting in low firm density. In the firm-level analysis, the main outcome is the number of months in a year that the firm operates. Again, firms exposed to frequent outages often shut down during outage periods relative to firms with reliable supply. Further, in the district-level analysis, U nreliableSupplyit is measured by the share of firms in a district citing electricity outages as a major constraint to their activities. In the 37 firm level analysis,51 U nreliableSupplyit , is a dummy variable equal to 1 if a firm reports electricity outages as a major constraint to its operation and 0 if otherwise. Arguably, ex- posure to outages is non-random across space and time, hence estimates from equation ˆ shows the association (9) cannot be ascribed causal interpretations. In other words, ϕ between exposure to outages and net entry of firms. Finally, I also explore how the asso- ciation between outages and (net) firm entry varies between ”high energy-intensive” and ”low energy-intensive” industries. Standard errors are clustered at the location level. Results Table 7 presents the results on the association between exposure to outages and firm den- sity, and extent of operation. Starting with the district-level analysis, column (1-2) shows the relationship between outage intensity and the number of firms operating in a district. According to the results, a 1 pp increase in the share of firms experiencing outages is as- sociated with a 1.7% reduction in the number of firms operating in the district.52 In other words, moving from a district with (fully) reliable supply of electricity to one with an un- reliable supply reduces the number of firms operating in the district by almost one-fifth. Obviously, proximity to markets is a key determinant of firm location (concentration). As a result, more (manufacturing) firms are likely to be located in densely populated areas like large cities relative to small towns.53 To account for this, in column 3-4, I use firm density (i.e., number of firms per 1000 people) as the outcome and explore the associa- tion with intensity of outages experienced by firms. Again the results show a negative association. These results suggest that unreliable electricity provision has negative im- pacts on the entry (and exit) of firms across locations. To further understand the effects of 51 Despite being an annual survey, information on unique firm identifiers in the LMMIS are scant (Hjort and Poulsen, 2019; Abebe et al., 2018). This limits the ability to exploit the panel structure of firms. Thus, in this analysis, I treat the data as a repeated cross-section rather than estimating within-firm variations. 52 ˆ expβ − 1 = exp−.0191 − 1 ≈ 0.17 53 Access to infrastructure like roads, electricity, water, and internet are also more readily available in densely populated areas cities relative to small towns 38 unreliable electricity provision firm entry, columns 5-6 and 6-7 present the results on the association between outage intensity and the number of (manufacturing) firms operating in ”high energy-intensive” and ”low energy-intensive” sectors respectively. Interestingly, the results show a negative and statistically significant effect of outages on the number of firms operating in ”high energy-intensive” sectors. Meanwhile, the effect on the num- ber of firms operating in ”low energy-intensive” sectors is also negative, albeit statistically insignificant. It is important to emphasize that the number (density) of firms operating each year is a net measure of firm entry and exit. While, outages are likely to affect both entry and exit, disentangling the effects is difficult due to data constraints. To provide insights into this, I conduct a firm-level analysis to explore the effect of firms’ exposure to outages and the duration of their operations in a given year. Given the importance of access to electricity in the production process, firms exposed to outages are more likely to shut down produc- tion during outage periods. Continuous shutdown of operations may contribute to their exit from the market. In panel B of Table 7, I show that firms with reliable supply op- erate for longer periods in a year than their counterparts with unreliable supply. Across the various specifications, the results suggest that conditional on the location, industry, ownership and time fixed effects, the average number of operating months is about one month lower for firms with unreliable electricity supply relative to their counterparts with reliable electricity supply. Interestingly, the effect holds for firms operating in both ”high energy-intensive” and ”low energy-intensive” sectors: perhaps an indication of the direct and indirect effects of outages on firm operations. While these estimates measure associ- ation, they provide suggestive evidence of the effects of outages on the entry and exit of firms in Ethiopia. 39 5.2.2 FDI Next, I explore the effects of unreliable electricity provision on the entry of foreign firms via FDI. In many developing and emerging countries, FDI is a potent source of firm entry as foreign investors enter these markets to harness the natural and human capital, and establish businesses in these economies. These new firms ultimately generate jobs for local ezina, 2021). For instance, Toews and V´ people (Toews and V´ ezina (2021), using data on Mozambique estimates to estimate the FDI-local job multiplier finds that each new FDI-job creates an additional 4.4 jobs of which 2.1 are formal jobs. In other words, aside the direct jobs creation associated with FDI’s (eg., employment of factory workers), FDI’s generates indirect jobs resulting from sectoral linkages.54 Thus, FDI is a key channel for job creation in many developing countries, hence, factors that constrain FDI flows could have negative implications on employment. The economic effects (including employment) associated with FDIs have made the at- traction of FDI a key development strategy of emerging economies. However, the quality of infrastructure provision such as electricity plays an important role in the flow (direc- tion) of FDI to developing economies. Cost of doing business is an important driver of FDI flows as it influence the returns on investment to investors. Meanwhile, the reliabil- ity of electricity provision is a significant determinant of the cost of doing business. For instance, during blackouts some firms rely on in-house electricity generation which are expensive relative to grid supply, while others curtail production. These options invari- ably have negative effects on firms’ profitability. In addition, persistent reliability issues in the electricity sector can generate adverse macroeconomic shocks in the economy thereby affecting investor confidence. In section, I ask: to what extent does unreliable provision of electricity affect FDI flows? To address this question, I exploit the ”Dumsor” power crisis in Ghana as a natural exper- 54 For instance, FDI into manufacturing, may generate increased demand for intermediate inputs within the value chain, which ultimately creates jobs. 40 iment and show how FDI flows to the country were affected by the crisis.55 Specifically, I focus on greenfield FDI in the non-energy-and-construction sectors (hereafter referred to as FDI [excl. energy & construction] projects) as: (i) the crisis could have induced signif- icant investment into the energy sector as part of efforts to resolve the crisis. Hence,FDI to the sector may rise in response to the crisis; and (ii) construction related FDIs, such as investment in real estate, are unlikely to respond to energy crisis as these sectors are relatively low energy intensive. To causally estimate the effect of the crisis on FDI in Ghana, I use the synthetic control approach (SCM) (Abadie and Gardeazabal, 2003; Abadie et al., 2010, 2011, 2015; Peri and Yasenov, 2019; Andersson, 2019). The basic idea behind the SCM is to construct a counter- factual, otherwise referred to as “synthetic control” using a weighted combination of the outcome variable in the control (donor) units that matches the outcome variable of the treated unit in the pre-treatment period. The weights are constructed via a data-driven al- gorithm by minimizing the differences in the outcome variable as well as key predictors of the outcome variable between the treated and control units. These weights are then used to construct a weighted combination of the outcome variable in the post-treatment period, which serves as a valid counterfactual for the treated unit during the post-treatment pe- riod. Therefore, the difference in the observed outcome between the treated unit and the synthetic control (counterfactual) measures the effect of the treatment.56 A unique fea- ture of the SCM is that, unlike the DiD, the so-called “parallel trends” assumption is not required for identification. The analysis is undertaken using a panel data on the greenfield FDI projects (excl. energy and construction sectors) in 23 emerging markets in Africa, Latin American and 55 There are at least two factors that motivate the choice of the Ghanaian crisis as a unique experiment: First, the crisis was a large scale shock to the economy as almost every economic sector and geographic administration in the country was affected. Second, the prolonged nature of the crisis (2013-2016) provides a unique opportunity to assess the effects of a relatively “long-term” shock as opposed to an instantaneous shock. 56 The SCM is somewhat similar to the DiD method, except that unlike the latter, the former does not impose equal weights on the control units 41 Caribbean, and Asia57 between 2007 and 2017. The choice of these countries was mainly informed by data availability and also countries that provide the relevant weights in con- structing the synthetic control.58 The dataset among others provides information on the number unique of FDI projects in each country, sector, and year. I focus mainly on count of investments rather than the monetary value of the investments as these FDI projects are heterogeneous across industry, country and year. Also the number of FDI projects into a country signals the response of various investors to the business environment in the destination country. The main predictors used include lags of the number of FDI [excl. energy & construction] projects, the log of GDP (2007-2008) and total FDI (all sectors) (2009-2012), GDP growth rate (2012) and population. Results: Ghana vs Synthetic Ghana The intuition behind the SCM is that if “Synthetic Ghana” is able to track the flow of FDI to Ghana in the pre-crisis era, then it provides a credible counterfactual of the flow of FDI to Ghana in the absence of the power crisis between 2013 and 2016. This section presents the main findings of the SCM. Tables A9 in the online appendix present the summary statistics of the predictors between Ghana vs. “Synthetic Ghana”, as well as the sample average during the pre-treatment period. Also, Table A10 in the online appendix shows the distribution of the weights used in constructing the “synthetic control” across the con- trol (donor) countries. Kenya, Zambia and Zimbabwe account for the largest share with weights of 0.57, 0.24, and 0.13 respectively. In terms of the main results, Figure 4 plots the number of FDI [excl. energy & construc- tion] projects in Ghana and Synthetic Ghana before, during and after the crisis. As shown 57 Cambodia, Ghana, Guatemala, Honduras, India, Jamaica, Kenya, Morocco, Mexico, Myanmar, Mau- ritius, Namibia, Nicaragua, Pakistan, Philippines, Senegal, South Africa, Uganda, Uruguay, Uzbekistan, Vietnam, Zambia, Zimbabwe 58 More so non-African countries were included in the control to limit the potential violation of the stable unit treatment value assumption (SUVTA) which is likely to be violated if, for instance, the power crisis in Ghana shifted FDI into neighboring countries or other African countries with similar business environment with the pre-crisis Ghana. 42 in the figure, prior to the crisis, FDI in synthetic Ghana tracks FDI in Ghana very closely. The average (absolute) difference in the number of FDI projects between Ghana and its synthetic counterpart in the pre-crisis period is just -0.34. However, as seen in the figure, the path-plot of the two units begins to diverge during the crisis period (2013-2016) and even persists in the post-crisis period (2017). Interestingly, the results show a consistent fall in the number of FDI [excl. energy & construction] projects in Ghana between 2013 and 2016, rising marginally in 2017. The gap (difference) in FDI between Ghana and its counterfactual in the pre/post pe- riod are shown in Figure 5. In 2014 (a year into the crisis), the gap between the number of non-energy FDI projects in Ghana and its synthetic counterpart is marginal (8%). How- ever, in 2015 (two years into the crisis), we observe a significant dip in FDI flows to the non-energy sectors in Ghana relative to its synthetic counterpart. For instance, in 2015 a total of 28 FDI projects in the non-energy-and-construction sectors were implemented in Ghana compared to an estimated counterfactual of 49.3. This represents 43% reduction in the number of FDI projects in the non-energy-and-construction sectors for the year. There was a slight improvement in 2016 with the gap reducing to 12%. The negative effects of the crisis persisted in 2017 – a year after the crisis was officially declared over, with a 27% reduction in the number of non-energy FDI project inflows. Overall, how did the power crisis affect FDI inflows? The estimates from the synthetic control analysis indicate that between 2013 and 2016, the power crisis on average resulted in a reduction in the number of non-energy sector FDI by 12.3% per annum. To assess the statistical significance of the gap between the observed FDI in Ghana and its synthetic counterpart, I follow the approach of Abadie et al. (2010, 2015) in construct- ing an “exact p-value”. The approach involves estimating iteratively the SCM for each country in the donor pool (i.e., the control countries) and obtaining the distribution of the respective placebo effects. Based on these placebo effects, the ratio of the root mean square prediction error (RMSPE) in the post-treatment period to the RMSPE for the pre- 43 treatment is computed and ranked across countries. The p-value is computed by dividing the rank of the treatment country (unit) by the total number of countries in the pool (Cun- ningham, 2021). Figure A5 in the appendix shows the distribution of the post-pre RSMPE ratio for all countries in the pool. Based on the results in the figure, Ghana is ranked second out of the 23 country units, which implies an “exact p-value” of 0.086.59 In other words, the SCM results are statistically significant at 10% error level. These results provide suggestive evidence that the quality of electricity supply is a ma- jor determinant of FDI flows. To the extent that unreliable provision of electricity supply reduces FDI inflows, the findings suggest that a reduction in firm entry via falling (green- field) FDIs is a mechanism through which electricity shortages affect employment since FDI is an important avenue for job creation in many developing countries (Toews and ezina, 2021). It is important to acknowledge that the estimated effect of outages on FDI V´ only reflects the impact on the entry of foreign firms. Robustness To test the robustness of the SCM results, I conduct two placebo (“in-time” and “in- space”) tests. Starting with the placebo “in-time” test, I explore the possibility of find- ing a similar gap in the FDI trends between Ghana and its synthetic control assuming a placebo treatment period. Specifically, I focus on the period 2007-2013 and assign 2011 as the treatment year. Finding a large placebo gap suggests that the results in Figures 4 and 5 are unlikely to be the causal impact of the power crisis on FDI. That is the effect could arise out of chance and not related to the impact of the power crisis. As shown in Figure A6, I find no evidence of significant divergence between Ghana and its synthetic control using the placement treatment period. In the case of the “in-space” placebo test, I iteratively assign treatment status to coun- tries in the control (donor) pool while using the remaining countries (including Ghana) 59 i.e., 1/23 = 0.086 44 to construct the respective counterfactual via the synthetic control method. The estimated gap in observed FDI and the counterfactual for the respective countries are compared to determine if the gap for Ghana is unusually large or comparable to the results for the other countries. Figure A7 (Panel A) shows the results of the placebo “in-space” test for all countries in the dataset. However, as shown in the plot, the SCM is unable to find a convex combination of donor units that replicates the gap plot for Ghana in the pre-treatment pe- riod. As a result, there is a large variance in the gap between Ghana and several countries in the pre-treatment period. Therefore following Abadie et al. (2010), in Panel B of A7, I drop countries where the pretreatment MSPE is considerably large (>20) than Ghana’s MSPE. Encouragingly aside from having a pre-treatment gap similar to Ghana, the results in Panel B shows that Ghana has the largest reduction in FDI in the post-treatment period. These placebo tests, together with the estimated p-value provide suggestive evidence that the SCM results in Figures 4 and 5 are statistically different from zero. In addition to the above, I also conduct additional robustness checks by performing the SCM analysis on FDI in the construction and energy sectors. Recall that up to this point the SCM analysis relies on data on FDI in sectors excluding energy and construction. In Figures A8 for instance, the path plot shows a slight increase in FDI into Ghana’s energy sector in Ghana relative counterfactual levels in 2014 and 2015, plausibly indicative of the Government’s response in attracting FDI into the sector to address the power deficit in the country. FDI in the sector, however, fell significantly relative to the counterfactual 2016 when the crisis ended. In the case of the construction sector, as shown in Figure A9, I do not find any consistent evidence of the effect of the power crisis on FDI in the sector. Between 2013 and 2014, construction FDI in Ghana increased relative to the counterfactual, however, we observe a complete reversal in 2015 and 2016. Thus, the net effect of the gap between FDI in Ghana and its counterfactual (synthetic Ghana) during the period of the power crisis is approximately zero. Overall, in line with the results in Section 5.2.1, the results herein provide additional 45 suggestive evidence of the negative impact of unreliable electricity provision on the entry of new firms in an economy. 6 Discussion and Conclusion Many African countries are confronted with rising unemployment. At the same time, many economies in the region grapple with unreliable provision of electricity to house- holds and industry, with potential implications on economic performance. The extent to which unreliable provision of electricity contributes to the growing unemployment in the region is a question central to development policy in the region. This paper presents evidence of how electricity shortages constrain job creation in Africa. To this end, I assemble an array of household and firm data across several SSA countries, along with unique data on greenfield FDI projects to: (i) estimate the extent to which outages affect employment, and (ii) the channels through which these impacts arise. Findings from the paper indicate that electricity outages are a major contributory factor to the growing unemployment in Africa. Skilled jobs and employment in non- agricultural sectors are the most affected. Specifically, from the cross-country analysis, the results from an IV regression estimation suggest that outages are associated with a 13.5 pp decrease in the probability of employment. Estimates from country-case studies in Ghana and Nigeria suggest that outages are associated with a 4.7 pp and 5.7 pp reduction in the probability of employment respectively. Factors such as cross-country heterogene- ity and differences in the measure of employment and outages across the various datasets could account for the differences in the point estimates. Nonetheless, the estimates fall in a similar ballpark, suggesting that outages reduce employment by between 4.7 and 13.5 pp. Further, the paper provides suggestive evidence of two plausible mechanisms. First outages constrain firm entry via: a reduction in firm density, and foreign direct invest- 46 ment in non-energy-and-construction sectors due to the effect of outages on the cost of doing business thereby serving as a disincentive to entrepreneurs and investors. Secondly, outages have a significant negative effect on the performance of incumbent firms thereby constraining the expansion of the productive sectors of the economy. The findings of this paper have important implications for policy. The results provide suggestive evidence that resolving the challenges in the electricity sector is an important channel towards expanding the industrial sector as it will increase job creation. Again, access to reliable electricity is crucial for improving the productivity of incumbent firms and attracting new (domestic and foreign) firms including FDI. These will contribute to increasing the number of employment opportunities in the region and ultimately reduce the high rate of unemployment in many African economies. 47 References Abadie, A., Diamond, A., and Hainmueller, J. (2010). Synthetic control methods for com- parative case studies: Estimating the effect of California’s tobacco control program. Jour- nal of the American Statistical Association, 105(490):493–505. Abadie, A., Diamond, A., and Hainmueller, J. (2011). Synth: An R package for synthetic control methods in comparative case studies. Journal of Statistical Software, 42(13). Abadie, A., Diamond, A., and Hainmueller, J. (2015). Comparative politics and the syn- thetic control method. American Journal of Political Science, 59(2):495–510. Abadie, A. and Gardeazabal, J. (2003). The economic costs of conflict: A case study of the Basque country. American Economic Review, 93(1):113–132. Abbasi, M., Lebrand, M. S. M., Mongoue, A. B., Pongou, R., Zhang, F., et al. (2022). Roads, electricity, and jobs: Evidence of infrastructure complementarity in sub-saharan africa. World Bank. Abebe, G., McMillan, M. S., and Serafinelli, M. (2018). Foreign direct investment and knowledge diffusion in poor locations. Technical report, National Bureau of Economic Research. Abeberese, A. B. (2017). Electricity cost and firm performance: Evidence from india. Re- view of Economics and Statistics, 99(5):839–852. Abeberese, A. B., Ackah, C. G., and Asuming, P. O. (2021). Productivity losses and firm re- sponses to electricity shortages: Evidence from Ghana. The World Bank Economic Review, 35(1):1–18. Adepitan, J. and Oladiran, E. (2012). Analysis of the dependence of power outages on 48 lightning events within the Ijebu province, Nigeria. Research Journal of Environmental and Earth Sciences, 4(9):850–856. Akpandjar, G. and Kitchens, C. (2017). From darkness to light: The effect of electrification in Ghana, 2000–2010. Economic Development and Cultural Change, 66(1):000–000. Alam, M. (2013). Coping with blackouts: Power outages and firm choices. Department of Economics, Yale University. Allcott, H., Collard-Wexler, A., and O’Connell, S. D. (2016). How do electricity shortages affect productivity? Evidence from India. American Economic Review, 106(3):587–624. Andersen, T. B., Bentzen, J., Dalgaard, C.-J., and Selaya, P. (2011). Does the internet reduce corruption? Evidence from US states and across countries. The World Bank Economic Review. Andersen, T. B., Bentzen, J., Dalgaard, C.-J., and Selaya, P. (2012). Lightning, IT diffusion, and economic growth across US states. Review of Economics and Statistics, 94(4):903–924. Andersen, T. B. and Dalgaard, C.-J. (2013). Power outages and economic growth in Africa. Energy Economics, 38:19–23. Andersson, J. J. (2019). Carbon taxes and CO2 emissions: Sweden as a case study. American Economic Journal: Economic Policy, 11(4):1–30. Angrist, J. D. and Pischke, J.-S. (2009). Mostly harmless econometrics: An empiricist’s com- panion. Princeton university press. ¸a Assunc ˜ o, J., Lipscomb, M., Mobarak, A. M., and Szerman, D. (2014). Electrification, agricultural productivity and deforestation in Brazil. Technical report, Discussion paper. 49 Bahia, K., Castells, P., Cruz, G., Pedros, X., Pfutze, T., Rodriguez Castelan, C., and Winkler, H. (2020). The welfare effects of mobile broadband internet: Evidence from nigeria. World Bank Policy Research Working Paper, (9230). Bertoni, E., Di Maio, M., Molini, V., and Nistico, R. (2019). Education is forbidden: The effect of the boko haram conflict on education in North-East Nigeria. Journal of Develop- ment Economics, 141:102249. Blimpo, M. P. and Cosgrove-Davies, M. (2019). Electricity access in Sub-Saharan Africa: Uptake, reliability, and complementary factors for economic impact. The World Bank. Chakravorty, U., Pelli, M., and Marchand, B. U. (2014). Does the quality of electricity matter? Evidence from rural India. Journal of Economic Behavior & Organization, 107, Part A:228 – 247. Chisholm, W. and Cummins, K. (2006). On the use of LIS/OTD flash density in electric utility reliability analysis. In The Lightning Imaging Sensor InternationalWorkshop. ´ A., and Lugauer, S. (2010). The effect of household appliances on Coen-Pirani, D., Leon, female labor force participation: Evidence from microdata. Labour Economics, 17(3):503– 513. Cole, M. A., Elliott, R. J., Occhiali, G., and Strobl, E. (2018). Power outages and firm per- formance in Sub-Saharan Africa. Journal of Development Economics, 134:150 – 159. Cunningham, S. (2021). Yale University Press. Dewan, A., Ongee, E. T., Rafiuddin, M., Rahman, M. M., and Mahmood, R. (2018). Light- ning activity associated with precipitation and CAPE over Bangladesh. International Journal of Climatology, 38(4):1649–1660. Dinkelman, T. (2011). The effects of rural electrification on employment: New evidence from South Africa. American Economic Review, 101(7):3078–3108. 50 Dzansi, J., Puller, S., Street, B., and Yebuah-Dwamena, B. (2018). The vicious circle of blackouts and revenue collection in developing economies: Evidence from Ghana. Tech- nical report, Mimeo. Essers, D., Megersa, K., and Sanfilippo, M. (2021). The productivity gaps of female-owned firms: Evidence from ethiopian census data. Economic Development and Cultural Change, 69(2):645–683. Fisher-Vanden, K., Mansur, E. T., and Wang, Q. J. (2015). Electricity shortages and firm productivity: Evidence from China’s industrial firms. Journal of Development Economics, 114:172–188. Fried, S. and Lagakos, D. (2021). Rural electrification, migration and structural transfor- mation: Evidence from ethiopia. Regional Science and Urban Economics, page 103625. Garicano, L., Lelarge, C., and Van Reenen, J. (2016). Firm size distortions and the produc- tivity distribution: Evidence from France. American Economic Review, 106(11):3439–79. Greenwood, J., Seshadri, A., and Yorukoglu, M. (2005). Engines of liberation. The Review of Economic Studies, 72(1):109–133. Guriev, S., Melnikov, N., and Zhuravskaya, E. (2020). 3G internet and confidence in gov- ernment. Forthcoming, Quarterely Journal of Economics. Hardy, M. and McCasland, J. (2019). Lights Off, Lights On: The Effects of Electricity Shortages on Small Firms. The World Bank Economic Review, 35(1):19–33. Hjort, J. and Poulsen, J. (2019). The arrival of fast internet and employment in Africa. American Economic Review, 109(3):1032–79. Hsieh, C.-T. and Klenow, P. J. (2009). Misallocation and manufacturing TFP in China and India. The Quarterly Journal of Economics, 124(4):1403–1448. 51 IEA (2019). World Energy Outlook. International Energy Agency. Kuka, E., Shenhav, N., and Shih, K. (2020). Do human capital decisions respond to the returns to education? Evidence from DACA. American Economic Journal: Economic Policy, 12(1):293–324. Lewis, J. (2014). Short-run and long-run effects of household electrification. In Economic History Workshop, Queen’s University, Kingston, Ontario, Canada. Lewis, J. and Severnini, E. (2014). The value of rural electricity: Evidence from the rollout of the US power grid. Technical report, Technical report. Lipscomb, M., Mobarak, M. A., and Barham, T. (2013). Development effects of electrifica- tion: Evidence from the topographic placement of hydropower plants in Brazil. Ameri- can Economic Journal: Applied Economics, 5(2):200–231. Manacorda, M. and Tesei, A. (2020). Liberation technology: mobile phones and political mobilization in Africa. Econometrica, 88(2):533–567. ıguez-Castel´ Masaki, T., Ochoa, R., and Rodr´ an, C. (2020). Broadband internet and house- hold welfare in senegal. Mswane, L. and Gaunt, C. (2005). Lightning performance improvement of the Swaziland electricity board transmission system (66 kv and 132 kv lines)-results of the pilot project. In Power Engineering Society Inaugural Conference and Exposition in Africa. Peri, G. and Yasenov, V. (2019). The labor market effects of a refugee wave synthetic control method meets the mariel boatlift. Journal of Human Resources, 54(2):267–309. Ramey, V. A. and Francis, N. (2009). A century of work and leisure. American Economic Journal: Macroeconomics, 1(2):189–224. 52 Reinikka, R. and Svensson, J. (2002). Coping with poor public capital. Journal of Develop- ment Economics, 69(1):51–69. Restuccia, D. and Rogerson, R. (2008). Policy distortions and aggregate productivity with heterogeneous establishments. Review of Economic Dynamics, 11(4):707 – 720. Romps, D. M., Seeley, J. T., Vollaro, D., and Molinari, J. (2014). Projected increase in light- ning strikes in the United States due to global warming. Science, 346(6211):851–854. Steinbuks, J. and Foster, V. (2010). When do firms generate? Evidence on in-house elec- tricity supply in Africa. Energy Economics, 32(3):505 – 514. Stock, J. H. and Yogo, M. (2005). Testing for weak instruments in linear iv regression. Identification and inference for econometric models: Essays in honor of Thomas Rothenberg. ezina, P.-L. (2021). Resource discoveries, FDI bonanzas, and local multi- Toews, G. and V´ pliers: Evidence from Mozambique. The Review of Economics and Statistics, pages 1–36. Verner, E. and Gyongy ¨ ¨ G. (2020). Household debt revaluation and the real economy: osi, Evidence from a foreign currency debt crisis. American Economic Review, 110(9):2667– 2702. 53 Figures Figure 1: Relationship between Outages and Lightning Intensity The (binned) scatter plots above shows the relationship between outages and lightning intensity. The left panel shows the correlation between the share of communities experiencing outages and lightning intensity; while the right panel shows the correlation between the share of households in a community experiencing outages and lightning intensity. Figure 2: Electricity Rationing Schedule during the Power Crisis in Ghana (2015) This figure shows an example of the power rationing schedule by the Electricity Company of Ghana (ECG) during the ”dumsor” power crisis in 2015. Source: Daily Graphic (2015) 54 Figure 3: Event Study- The ”Dumsor” Power Crises and Unemployment in Ghana This figure shows the coefficients and 95% confidence intervals from an event study regression that estimate the interactions between year and high exposure indicators, where the outcome variable is an indicator for the employment status of a person, and year 2006 is the omitted category. High Exposure is an indicator variable equal to 1 if the district’s electricity access rate in the year 2000 is above the median access rate in the country and 0 if otherwise. The regression controls for gender, education, rural-urban status, access to roads, access to water, and the following fixed effects: district, year and birth-year. 55 Figure 4: Path Plot of Number of FDI Projects during 2007-2017: Ghana vs Synthetic Ghana This plot shows the trend in the number of ”non-energy” sector FDI projects in Ghana vs a Synthetic Ghana before and during the power crisis 56 Figure 5: Gap in the Number of FDI Projects between Ghana and Synthetic Ghana This plot shows the gap in the number of ”non-energy” sector FDI projects in Ghana and a Synthetic Ghana before and during the power crisis 57 Tables Table 1: First Stage Regression: Electricity Outages and Lightning Intensity Outages in Community (0/1) Outages in Community (% HHs) (1) (2) (3) (4) Lightning intensity (log) 0.119∗∗∗ 0.119∗∗∗ 0.087∗∗∗ 0.088∗∗∗ (0.012) (0.012) (0.008) (0.008) Controls Yes Yes Yes Yes Country FE Yes Yes Yes Yes Survey Year FE Yes No Yes No Survey Round FE No Yes No Yes Kleibergen-Paap F statistic 92.383 93.529 121.927 124.648 Observations 24999 24999 24999 24999 Notes: In columns 1-2, the dependent variable is a dummy set equal to 1 if more than 50% of connected households in the PSU do not have access to reliable electricity, and 0 if otherwise. In column 3-4, the dependent variable is the share of connected households in the PSU do not have access to reliable electricity. Controls included are gender, age, age squared, educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 58 Table 2: Electricity Shortages and Employment All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS Outages in Community (0/1) -0.021∗∗ -0.020∗∗ -0.028∗∗∗ -0.027∗∗∗ 0.007 0.007 -0.013 -0.012 -0.061∗∗∗ -0.060∗∗∗ (0.010) (0.010) (0.010) (0.010) (0.005) (0.005) (0.012) (0.012) (0.021) (0.021) IV Outages in Community (0/1) -0.137∗∗ -0.135∗∗ -0.137∗∗ -0.135∗∗ 0.000 0.000 -0.192∗∗∗ -0.191∗∗∗ -0.024 -0.020 (0.062) (0.062) (0.063) (0.063) (0.027) (0.027) (0.070) (0.070) (0.133) (0.133) Reduced Form Lightning Intensity (log) -0.016∗∗ -0.016∗∗ -0.016∗∗ -0.016∗∗ 0.000 0.000 -0.024∗∗∗ -0.024∗∗∗ -0.003 -0.002 (0.007) (0.007) (0.007) (0.008) (0.003) (0.003) (0.008) (0.008) (0.016) (0.016) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Mean dep. Var 0.583 0.583 0.522 0.522 0.061 0.061 0.658 0.658 0.486 0.486 Kleibergen-Paap F statistic 92.383 93.529 92.383 93.529 92.383 93.529 95.539 96.389 43.303 43.749 Observations 24999 24999 24999 24999 24999 24999 16924 16924 4143 4143 Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are gender, age (log) and educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 59 Table 3: Effect of Power Crises on Employment in Ghana GLSS Afrobarometer (1) (2) (3) (4) (5) (6) High Exposure × Power Crisis -0.0455*** -0.0368** -0.0341** -0.0515* -0.0499* -0.0472* (0.0124) (0.0175) (0.0171) (0.0268) (0.0267) (0.0283) District FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes YOB FE Yes Yes Yes Yes Yes Yes Indiv. & Comm. Controls No Yes Yes No Yes Yes District Ctrls × Trend No No Yes No No Yes Mean dep. var 0.7399 0.7474 0.7474 0.7271 0.7274 0.7274 R-squared 0.3632 0.3812 0.3822 0.1371 0.1434 0.1435 Survey Rounds 4 4 4 6 6 6 Observations 102487 41464 41464 8663 8644 8644 Notes: Dependent variable is a dummy variable equal to 1 if the individual is employed and 0 if otherwise. Indiv. & Comm. Controls represent individual and community attributes such as gender, highest educational attainment of the respondent, rural/urban status, access to road and water in the community. District Ctrls × Trend include the baseline homeownership and literacy rates in the district interacted with a time trend. Standard errors are clustered at the district level. OLS estimations. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 60 Table 4: Panel Fixed Effect Regression: Electricity outages and Employment in Nigeria Employed (1/0) (1) (2) (3) (4) (5) (6) Outages in Community (0/1) -0.0565** -0.0585* -0.0631** (0.0277) (0.0306) (0.0290) Frequency of outages Daily -0.0519* -0.0540* -0.0605** (0.0290) (0.0317) (0.0298) Several times a week -0.0569** -0.0604* -0.0627** (0.0285) (0.0322) (0.0310) Several times a month -0.0826*** -0.0841*** -0.0831** (0.0303) (0.0321) (0.0325) Several times a year -0.0550 -0.0596 -0.0524 (0.0385) (0.0433) (0.0439) Indiv Controls No Yes Yes No Yes Yes HH FE Yes Yes Yes Yes Yes Yes State× Year FE Yes Yes Yes Yes Yes Yes Excluding North East No No Yes No No Yes Mean dep. var 0.7850 0.7883 0.7937 0.7850 0.7883 0.7937 R-squared 0.3081 0.3450 0.3358 0.3083 0.3452 0.3360 Survey Rounds 3 3 3 3 3 3 Observations 16134 12888 11907 16134 12888 11907 Notes: Dependent variable is a dummy equal to 1 if the individual is employed and 0 if otherwise. Individual Controls included gender, educational attainment and rural/urban status. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 61 Table 5: IV Regression: Electricity outages and Firm Performance Dependent Var: Log Sales Log Sales/Worker Log Value Added/Worker (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) OLS Outages (log) -0.1441∗ -0.1427∗ -0.1137 -0.1137 0.0388 0.0388 (0.0739) (0.0754) (0.0753) (0.0747) (0.0903) (0.0903) Outage Hours (log) -0.1625∗∗∗ -0.1554∗∗∗ -0.1242∗∗∗ -0.1178∗∗∗ -0.0266 -0.0266 (0.0432) (0.0426) (0.0318) (0.0313) (0.0442) (0.0442) IV Outages (log) -1.1536∗ -1.1672∗∗ -1.2862∗∗ -1.2700∗∗ -2.3427∗∗ -2.3427∗∗ (0.5798) (0.5575) (0.6084) (0.5870) (0.8865) (0.8865) Outage Hours (log) -0.5989∗ -0.6132∗ -0.6850∗ -0.6828∗ -2.4713∗ -2.4713∗ (0.3271) (0.3212) (0.3535) (0.3472) (1.4199) (1.4199) Reduced Form Lightning intensity (log) -0.7650∗∗ -0.7856∗∗ -0.7228∗∗ -0.7446∗∗ -0.8856∗∗ -0.8873∗∗ -0.8554∗∗ -0.8548∗∗ -1.1596∗∗∗ -1.1596∗∗∗ -1.1708∗∗∗ -1.1708∗∗∗ (0.3533) (0.3441) (0.3532) (0.3429) (0.3532) (0.3422) (0.3592) (0.3451) (0.3577) (0.3577) (0.3476) (0.3476) Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Kleibergen-Paap F statistic 31.7013 30.5283 19.6054 19.9313 35.6018 34.4263 21.7683 21.6960 20.8382 20.8382 4.7931 4.7931 Observations 3789 3789 3617 3617 3743 3743 3572 3572 1616 1616 1556 1556 Notes: Firm Controls include: age of the firm, whether the firm is foreign or domestic, and the mobile phone coverage rate in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 62 Table 6: IV Regression: Electricity outages and Labor Demand First Stage Dep. Var: Log # of Full Time Workers Log # of Temp. Workers Log Labor Cost Log Labor Cost per Worker (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) OLS Outages (log) -0.0197 -0.0186 0.0409 0.0411 -0.2057∗∗ -0.2061∗∗ -0.1905∗∗∗ -0.1919∗∗∗ (0.0356) (0.0346) (0.0338) (0.0350) (0.0840) (0.0819) (0.0601) (0.0589) Outage Hours (log) -0.0227 -0.0232 0.0021 0.0031 -0.2095∗∗∗ -0.2057∗∗∗ -0.1784∗∗∗ -0.1745∗∗∗ (0.0244) (0.0239) (0.0238) (0.0261) (0.0580) (0.0542) (0.0444) (0.0417) IV Outages (log) -0.1654 -0.1780 -0.6165∗∗∗ -0.5809∗∗∗ -1.1320∗ -1.1335∗ -1.1494∗ -1.1499∗ (0.2358) (0.2459) (0.2056) (0.2142) (0.6505) (0.6084) (0.6777) (0.6348) ∗∗∗ ∗∗∗ ∗ Outage Hours (log) -0.0818 -0.0909 -0.3359 -0.3195 -0.6256 -0.6342 -0.6348 -0.6406∗ (0.1228) (0.1301) (0.1044) (0.1086) (0.3812) (0.3614) (0.4001) (0.3774) Reduced Form Lightning intensity (log) -0.1000 -0.1092 -0.0942 -0.1052 -0.3726∗∗∗ -0.3564∗∗ -0.3868∗∗∗ -0.3698∗∗ -0.7288∗ -0.7439∗ -0.7195∗ -0.7359∗ -0.7357∗ -0.7503∗∗ -0.7292∗ -0.7423∗∗ (0.1447) (0.1549) (0.1442) (0.1552) (0.1228) (0.1377) (0.1234) (0.1379) (0.3914) (0.3799) (0.3908) (0.3759) (0.3709) (0.3509) (0.3708) (0.3457) Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Kleibergen-Paap F statistic 25.6449 24.1860 22.8923 22.6933 25.6449 24.1860 22.8923 22.6933 31.3946 29.6312 16.9568 17.3329 30.7593 29.0558 16.9526 17.3191 Observations 4281 4281 4076 4076 4281 4281 4076 4076 3809 3809 3650 3650 3803 3803 3644 3644 63 Notes: Firm Controls include: age of the firm, whether the firm is foreign or domestic, and the mobile phone coverage in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level Table 7: Electricity Shortages and and Net Firm Entry (1) (2) (3) (4) (5) (6) (7) (8) A: District Level Analysis All Firms High Energy Intensive Low Energy Intensive log (# of firms) # firms per 1000 people log (# of firms) Share of firms with unreliable supply -0.192** -0.191** -0.013* -0.013* -0.133** -0.137** -0.091 -0.086 (0.08) (0.079) (0.008) (0.007) (0.061) (0.061) (0.060) (0.058) Controls No Yes No Yes No Yes No Yes District FE Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Yes Yes R2 0.857 0.858 0.955 0.964 939 939 939 939 MP 1.486 1.486 0.106 0.106 0.9687 0.9687 1.4456 1.4456 Obs 939 939 939 939 939 939 939 939 B: Firm Level Analysis All Firms High Energy Intensive Low Energy Intensive # of months operating Firm experiencing unreliable supply (0/1) -0.905*** -1.008*** -0.911*** -0.987*** -1.020*** -1.036*** -0.856*** -0.845*** (0.066) (0.063) (0.067) (0.065) (0.105) (0.107) (0.065) (0.065) Controls No Yes No Yes No No No No District FE Yes Yes No No Yes No Yes No Year FE Yes No Yes Yes Yes Yes Yes Yes City/Town FE Yes Yes Yes Yes Yes Yes Yes Yes Ownership type FE Yes No Yes Yes Yes Yes Yes Yes Industry FE No Yes No No No No No No IndustryXYear FE No No Yes Yes No Yes No Yes R2 0.171 0.196 0.199 0.207 0.108 0.142 0.205 0.238 MP 10.43 10.45 10.436 10.459 10.742 10.747 10.271 10.279 Obs 15881 11526 15814 11479 5401 5345 10453 10387 Notes: In the district level analysis, controls include mobile network penetration at the district level, baseline nightlight intensity and population interacted with time trends. In addition to these controls, the firm level analysis, include female ownership share at the firm level. Standard errors are clustered at district level (panel A) and city level (panel B). ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 64 A ONLINE APPENDIX A.1 Figures A.1.1 Figures on Lightning Instrument Figure A1: Correlation between Lightning Intensity and CAPE× Precip This plot shows the correlation between actual lightning strikes measured by NASA’s LIS/OTD satellite and CAPE× Precipitation rate. The y-axis measures the mean annual flash rate between 1995 ad 2010. The x-axis also measures the mean annual CAPE× Precipitation rate over the same period. 65 Figure A2: Actual Lightning Intensity and CAPE× Precipitation Rate The left panel shows the average mean annual lightning flash rate (km2 /yr) between 1995 ad 2010 measured by the NASA’s LIS/OTD satellite. The right panel shows the mean annual lightning intensity proxied by CAPE× Precipitation Rate (J kg−1 mm hr−1 ) over the same period. 66 A.1.2 Additional figures on household-level analysis Figure A3: Outages and Unemployment: Assessing the sensitivity of estimates to the threshold for defining outage dummy This Figure shows point estimates and 95% confidence interval. Each estimate corresponds to β in equation 3, where outage is defined as a dummy if the share of households in a locality experiencing outages is at least the number on the horizontal axis. 67 Figure A4: Leave-one-out: first-stage Relationship This Figure shows point estimates and 95% confidence interval of the first-stage relationship between out- ages and our measure of lightning intensity. Each estimate corresponds to ϕ in equation 2 estimated on the sample after excluding data from the respective countries shown on the x-axis. 68 A.1.3 Additional figures on the synthetic control analysis Figure A5: Ratio Test-Ratios of Post-treatment MSPE to Pre-treatment RMSPE: Ghana and 22 control countries Figure A6: Placebo in-Time Tests This plot shows the trends in FDI projects in Ghana and synthetic Ghana when a placebo crisis is imposed in 2010 (three years before the actual crisis) 69 Figure A7: Permutation Test: No. of FDI projects gaps in Ghana and Placebo Gaps for the control countries (a) Panel A (b) Panel B Panel A shows the number of FDI projects gap in Ghana and placebo gaps in all the 22 control countries. Panel B shows the number of FDI projects gap in Ghana and placebo gaps in the X control countries after excluding countries with a pre-treatment MSPE 5 times Ghana’s pre-treatment MSPE Figure A8: FDI in the Energy Sector: Ghana vs Synthetic Ghana This figure shows trends in the number of FDI projects in the energy sector in Ghana vs a Synthetic Ghana before and during the power crisis 70 Figure A9: FDI in the Construction Sector: Ghana vs Synthetic Ghana This figure shows trends in the number of FDI projects in the construction sector in Ghana vs a Synthetic Ghana before and during the power crisis 71 A.2 Tables A.2.1 Summary statistics of all datasets Table A1: Summary statistics Variable Mean Std. Dev. Min. Max. N Afrobarometer Employment 0.593 0.491 0 1 29440 Outages in Comm. 0.633 0.482 0 1 42776 Outages in Comm.(% HHs) 0.578 0.35 0 1 42776 No educ 0.081 0.273 0 1 42582 Informal 0.032 0.175 0 1 42582 Primary 0.206 0.404 0 1 42582 Secondary 0.473 0.499 0 1 42582 Tertiary 0.209 0.406 0 1 42582 Precipitation (log) 0.035 0.026 0 0.144 42711 Temperature (log) 3.132 0.175 2.278 3.438 42711 Age 36.528 14.674 18 99 42679 Female 0.503 0.5 0 1 42776 Mobile Network Coverage 0.872 0.236 0 1 36577 Lightning Intensity (log) 10.497 1.287 6.23 12.186 42711 WBES Sales (log) 11.575 2.206 7.072 15.913 5506 Sales per worker (log) 8.881 1.646 0.554 11.905 5440 Value-added per worker (log) 8.164 1.438 2.741 10.832 2231 # of Workers (log) 2.654 1.058 0 4.927 6271 Labor cost (log) 9.49 1.94 3.351 13.199 5531 Outages (log) 1.923 0.896 0 3.434 5258 Outage Hours (log) 3.104 1.42 0 6.443 4871 Lightning intensity (log) 11.058 0.669 9.636 12.783 7591 Precipitation (log) 0.037 0.014 0.018 0.099 7591 Temperature (log) 3.181 0.111 2.908 3.353 7591 Age of firm 14.415 10.726 0 41 7515 Foreign ownership 0.198 0.399 0 1 7489 Mobile Network Coverage 0.857 0.226 0 1 7193 Manufacturing 0.482 0.5 0 1 7591 Ghana: GLSS Employment 0.739 0.439 0 1 103135 HighExposure X PowerCrisis 0.142 0.349 0 1 104525 Female 0.539 0.498 0 1 105193 Rural 0.612 0.487 0 1 105193 Education No formal educ 0.224 0.417 0 1 70208 Basic 0.33 0.47 0 1 70208 Secondary 0.38 0.485 0 1 70208 Tertiary 0.066 0.248 0 1 70208 Access to road 0.856 0.351 0 1 69533 Nigeria: GHS Employment 0.763 0.425 0 1 33637 Female 0.533 0.499 0 1 41475 Education No formal educ 0.158 0.365 0 1 28230 Basic 0.351 0.477 0 1 28230 Secondary 0.334 0.471 0 1 28230 Tertiary 0.157 0.364 0 1 28230 Urban 0.302 0.459 0 1 41475 Outage 0.967 0.179 0 1 21117 Outage freq. Never 0.033 0.179 0 1 21117 Everyday 0.52 0.5 0 1 21117 Several times a week 0.303 0.459 0 1 21117 Several times a month 0.108 0.31 0 1 21117 Several times a year 0.037 0.189 0 1 21117 72 A.2.2 Additional results on household-level analysis Table A2: Sectoral Distribution of Impacts of Electricity Shortages on Employment Sector of Employment All Sectors Private Sector Public Sector All Private Firms Self Employ (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS Outages in Community (0/1) -0.021∗∗ -0.020∗∗ -0.023∗∗ -0.022∗∗ -0.014∗ -0.013∗ -0.009 -0.009 -0.001 -0.001 (0.010) (0.010) (0.010) (0.010) (0.008) (0.008) (0.008) (0.008) (0.006) (0.006) IV Outages in Community (0/1) -0.137∗∗ -0.135∗∗ -0.305∗∗∗ -0.302∗∗∗ -0.316∗∗∗ -0.312∗∗∗ 0.010 0.010 0.146∗∗∗ 0.145∗∗∗ (0.062) (0.062) (0.070) (0.070) (0.067) (0.067) (0.045) (0.045) (0.039) (0.039) Reduced Form Lightning Intensity (log) -0.016∗∗ -0.016∗∗ -0.036∗∗∗ -0.036∗∗∗ -0.037∗∗∗ -0.037∗∗∗ 0.001 0.001 0.017∗∗∗ 0.017∗∗∗ (0.007) (0.007) (0.008) (0.008) (0.007) (0.007) (0.005) (0.005) (0.004) (0.004) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Mean dep. Var 0.583 0.583 0.463 0.463 0.204 0.204 0.259 0.259 0.109 0.109 Kleibergen-Paap F statistic 92.383 93.529 90.489 91.756 90.489 91.756 90.489 91.756 90.489 91.756 Observations 24999 24999 24415 24415 24415 24415 24415 24415 24415 24415 Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are gender, age (log) and educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 73 Table A3: Electricity Shortages and Employment: Effects by Gender All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) IV Regression Male Outages in Community (0/1) -0.125∗ -0.123 -0.163∗∗ -0.160∗∗ 0.038 0.037 -0.221∗∗∗ -0.219∗∗∗ 0.179 0.182 (0.076) (0.075) (0.077) (0.076) (0.036) (0.035) (0.085) (0.085) (0.175) (0.173) Mean dep. Var 0.631 0.631 0.557 0.557 0.074 0.074 0.693 0.693 0.592 0.592 Fstat 89.015 89.758 89.015 89.758 89.015 89.758 80.764 81.300 40.903 41.698 Obs 13283 13283 13283 13283 13283 13283 9127 9127 1848 1848 Female Outages in Community (0/1) -0.159∗ -0.155∗ -0.109 -0.106 -0.050 -0.049 -0.176∗ -0.176∗ -0.206 -0.192 (0.086) (0.086) (0.087) (0.088) (0.035) (0.035) (0.098) (0.098) (0.183) (0.184) Mean dep. Var 0.527 0.527 0.482 0.482 0.046 0.046 0.618 0.618 0.400 0.400 Kleibergen-Paap F statistic 80.278 81.540 80.278 81.540 80.278 81.540 77.932 78.815 23.519 23.873 Observations 11716 11716 11716 11716 11716 11716 7797 7797 2295 2295 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are age (log) and educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level Table A4: Electricity Shortages and Employment using an Alternate Measure of Outages All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS Outages in Community (% HHs) -0.053∗∗∗ -0.051∗∗∗ -0.072∗∗∗ -0.069∗∗∗ 0.019∗∗ 0.018∗∗ -0.040∗∗ -0.040∗∗ -0.106∗∗∗ -0.103∗∗∗ (0.015) (0.015) (0.015) (0.015) (0.008) (0.008) (0.018) (0.018) (0.032) (0.032) IV Outages in Community (% HHs) -0.186∗∗ -0.182∗∗ -0.187∗∗ -0.183∗∗ 0.001 0.001 -0.266∗∗∗ -0.263∗∗∗ -0.032 -0.026 (0.084) (0.083) (0.085) (0.084) (0.037) (0.036) (0.095) (0.094) (0.174) (0.173) Reduced Form Lightning Intensity (log) -0.016∗∗ -0.016∗∗ -0.016∗∗ -0.016∗∗ 0.000 0.000 -0.024∗∗∗ -0.024∗∗∗ -0.003 -0.002 (0.007) (0.007) (0.007) (0.008) (0.003) (0.003) (0.008) (0.008) (0.016) (0.016) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Mean dep. Var 0.583 0.583 0.522 0.522 0.061 0.061 0.658 0.658 0.486 0.486 Kleibergen-Paap F statistic 121.927 124.648 121.927 124.648 121.927 124.648 119.875 122.001 61.209 61.692 Observations 24999 24999 24999 24999 24999 24999 16924 16924 4143 4143 Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are gender, age, age squared, educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 74 Table A5: First Stage Regression: Lightning Intensity and Outages at the District Level Dep. Var: Electricity Outages in District (0/1) All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Lightning intensity (log) 0.133∗∗∗ 0.134∗∗∗ 0.133∗∗∗ 0.134∗∗∗ 0.133∗∗∗ 0.134∗∗∗ 0.120∗∗∗ 0.122∗∗∗ 0.161∗∗∗ 0.162∗∗∗ (0.033) (0.034) (0.033) (0.034) (0.033) (0.034) (0.035) (0.035) (0.035) (0.035) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Kleibergen-Paap F statistic 15.684 15.959 15.684 15.959 15.684 15.959 12.148 12.197 20.720 21.019 Observations 31380 31380 31380 31380 31380 31380 21637 21637 5233 5233 Notes: Dependent variable is a is defined as 1 if more than 50% of connected households in the PSU do not have access to reliable electricity, and 0 if otherwise. Controls included are gender, age, age squared, educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level Table A6: Electricity Shortages and Employment using District Level Measures of Expo- sure to Outages All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS Outages in District (0/1) -0.011 -0.010 -0.018 -0.019 0.008 0.009 -0.008 -0.009 -0.013 -0.010 (0.012) (0.013) (0.015) (0.015) (0.012) (0.012) (0.015) (0.015) (0.021) (0.021) IV Outages in District (0/1) -0.157∗∗ -0.148∗∗ -0.181∗∗∗ -0.178∗∗∗ 0.024 0.030 -0.278∗∗∗ -0.263∗∗∗ -0.063 -0.061 (0.067) (0.066) (0.068) (0.068) (0.042) (0.043) (0.098) (0.095) (0.085) (0.084) Reduced Form Lightning Intensity (log) -0.021∗∗ -0.020∗∗ -0.024∗∗ -0.024∗∗ 0.003 0.004 -0.034∗∗∗ -0.032∗∗∗ -0.010 -0.010 (0.009) (0.009) (0.009) (0.010) (0.006) (0.006) (0.010) (0.011) (0.014) (0.014) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Mean dep. Var 0.568 0.568 0.474 0.474 0.094 0.094 0.636 0.636 0.469 0.469 Kleibergen-Paap F statistic 15.684 15.959 15.684 15.959 15.684 15.959 12.148 12.197 20.720 21.019 Observations 31380 31380 31380 31380 31380 31380 21637 21637 5233 5233 Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are gender, age (log) and educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 75 Table A7: Electricity Shortages and Employment: Controlling for Electrification All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS Outages in Community (0/1) -0.018∗ -0.017∗ -0.023∗∗ -0.022∗∗ 0.006 0.005 -0.008 -0.008 -0.059∗∗∗ -0.058∗∗∗ (0.010) (0.010) (0.010) (0.010) (0.005) (0.005) (0.011) (0.011) (0.021) (0.021) IV Outages in Community (0/1) -0.119∗ -0.117∗ -0.111∗ -0.109∗ -0.008 -0.008 -0.170∗∗ -0.168∗∗ -0.013 -0.008 (0.064) (0.064) (0.065) (0.064) (0.028) (0.028) (0.071) (0.071) (0.136) (0.136) Reduced Form Lightning Intensity (log) -0.014∗ -0.014∗ -0.013∗ -0.013∗ -0.001 -0.001 -0.021∗∗ -0.021∗∗ -0.001 -0.001 (0.007) (0.007) (0.008) (0.008) (0.003) (0.003) (0.008) (0.008) (0.016) (0.016) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Mean dep. Var 0.583 0.583 0.522 0.522 0.061 0.061 0.658 0.658 0.486 0.486 Kleibergen-Paap F statistic 88.741 89.856 88.741 89.856 88.741 89.856 91.412 92.240 42.152 42.641 Observations 24999 24999 24999 24999 24999 24999 16924 16924 4143 4143 Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are gender, age (log) and educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. In addition, we control for an electricity access dummy equal to 1 if the community access rate is above the median and 0 if otherwise. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 76 Table A8: Electricity Shortages and Employment: Restricting to places with universal ac- cess to electricity All Skilled Workers Unskilled Workers Employed (0/1) Employed in Non-Agric (0/1) Employed in Agric (0/1) Employed (0/1) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS Outages in Community (0/1) -0.015 -0.015 -0.024∗ -0.024∗ 0.009 0.008 -0.005 -0.005 -0.051 -0.052 (0.014) (0.014) (0.014) (0.014) (0.007) (0.007) (0.015) (0.015) (0.033) (0.033) IV Outages in Community (0/1) -0.114∗ -0.113∗ -0.110∗ -0.109∗ -0.004 -0.004 -0.162∗∗ -0.161∗∗ 0.030 0.028 (0.060) (0.060) (0.061) (0.061) (0.021) (0.021) (0.065) (0.065) (0.135) (0.135) Reduced Form Lightning Intensity (log) -0.017∗ -0.017∗ -0.017∗ -0.017∗ -0.001 -0.001 -0.026∗∗ -0.026∗∗ 0.004 0.004 (0.009) (0.009) (0.009) (0.010) (0.003) (0.003) (0.010) (0.010) (0.020) (0.020) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Survey Year FE Yes No Yes No Yes No Yes No Yes No Survey Round FE No Yes No Yes No Yes No Yes No Yes Mean dep. Var 0.632 0.632 0.582 0.582 0.050 0.050 0.715 0.715 0.528 0.528 Kleibergen-Paap F statistic 125.912 126.511 125.912 126.511 125.912 126.511 121.678 122.135 55.596 55.649 Observations 12607 12607 12607 12607 12607 12607 8718 8718 1871 1871 Notes: Dependent variable(s) is a measure of employment status of the individual. It is a dummy equal to 1 if the individual is employed and 0 if otherwise. Controls included are gender, age (log) and educational attainment, mobile phone coverage, and the logs of total precipitation and mean annual temperature. Standard errors clustered at the PSU level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 77 A.2.3 Additional Results on Synthetic Control Table A9: Mean of Predictors Before The Power Crisis in 2013 Variables Ghana Synth. Ghana Sample Avg. No. FDI [excl. energy & construction] projects (2008) 15 18.71 90.14 No. FDI [excl. energy & construction] projects (2009) 23 21.83 71.05 No. FDI [excl. energy & construction] projects (2011) 46 44.04 85.55 No. FDI [excl. energy & construction] projects (2012) 38 38.31 77.55 No. FDI [All sectors] projects (2009) 28 25.37 82 No. FDI [All sectors] projects (2010) 27 27.838 76.682 No. FDI [All sectors] projects (2011) 49 47.881 94.455 No. FDI [All sectors] projects (2012) 42 42.659 86.5 Log GDP (2007) 23.98 24.04 24.46 Log GDP (2008) 24.07 24.04 24.50 Log GDP (2012) 24.42 24.32 24.68 Population (mill) (2007-2012) 24.48 38.30 91.95 GDP growth (2012) 9.29 6.78 5.21 Table A10: Country Weights in Synthetic Ghana Country Weights Country Weights Guatemala 0 Nicaragua 0 Honduras 0 Pakistan 0.045 India 0.001 Philippines 0.002 Jamaica 0 Senegal 0 Kenya 0.571 Uganda 0 Cambodia 0.001 Uruguay 0 Morocco 0.001 Uzbekistan 0.002 Mexico 0.002 Vietnam 0.001 Myanmar 0 South Africa 0.006 Mauritius 0 Zambia 0.242 Namibia 0 Zimbabwe 0.127 78 A.2.4 Additional Results on firm-level analysis Table A11: First Stage Regression: Electricity Shortages and Firm Performance Log Sales Log Sales/Worker Log Value Added/Worker First Stage Dep. Var: Outages (log) Outage Hours (log) Outages (log) Outage Hours (log) Outages (log) Outage Hours (log) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Lightning intensity (log) 0.6631 ∗∗∗ 0.6730 ∗∗∗ 1.2069 ∗∗∗ 1.2142 ∗∗∗ 0.6886 ∗∗∗ 0.6986 ∗∗∗ 1.2488 ∗∗∗ 1.2518 ∗∗∗ 0.4950 ∗∗∗ 0.4950 ∗∗∗ 0.4738 ∗∗ 0.4738∗∗ (0.1178) (0.1218) (0.2726) (0.2720) (0.1154) (0.1191) (0.2677) (0.2687) (0.1084) (0.1084) (0.2164) (0.2164) Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Kleibergen-Paap F statistic 31.7013 30.5283 19.6054 19.9313 35.6018 34.4263 21.7683 21.6960 20.8382 20.8382 4.7931 4.7931 Observations 3789 3789 3617 3617 3743 3743 3572 3572 1616 1616 1556 1556 Notes: Firm Controls included are age of the firm, whether the firm is foreign or domestic, and mobile coverage rate in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 79 Table A12: First Stage Regression: Electricity Shortages and Labor Demand Log # of Full Time Workers Log # of Temp. Workers Log Labor Cost Log Labor Cost per Worker First Stage Dep. Var: Outages (log) Outage Hours (log) Outages (log) Outage Hours (log) Outages (log) Outage Hours (log) Outages (log) Outage Hours (log) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) Lightning intensity (log) 0.6043∗∗∗ 0.6135∗∗∗ 1.1516∗∗∗ 1.1575∗∗∗ 0.6043∗∗∗ 0.6135∗∗∗ 1.1516∗∗∗ 1.1575∗∗∗ 0.6438∗∗∗ 0.6563∗∗∗ 1.1501∗∗∗ 1.1603∗∗∗ 0.6400∗∗∗ 0.6525∗∗∗ 1.1488∗∗∗ 1.1589∗∗∗ (0.1193) (0.1247) (0.2407) (0.2430) (0.1193) (0.1247) (0.2407) (0.2430) (0.1149) (0.1206) (0.2793) (0.2787) (0.1154) (0.1211) (0.2790) (0.2785) Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Kleibergen-Paap F statistic 25.6449 24.1860 22.8923 22.6933 25.6449 24.1860 22.8923 22.6933 31.3946 29.6312 16.9568 17.3329 30.7593 29.0558 16.9526 17.3191 Observations 4281 4281 4076 4076 4281 4281 4076 4076 3809 3809 3650 3650 3803 3803 3644 3644 Notes: Firm Controls include: age of the firm, whether the firm is foreign or domestic, and mobile phone coverage rate in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 80 Table A13: Electricity outages and Firm Performance: Effects Across Firm Size Dependent Var: Log Sales Log Sales/Worker Log Value Added/Worker (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) IV Regression Small Firms Outages (log) -0.8995 -0.9083 -1.2239∗ -1.2039∗ -5.0250∗ -5.0250∗ (0.7149) (0.7072) (0.7200) (0.7049) (2.6130) (2.6130) Outage Hours (log) -0.4185 -0.4310 -0.5946 -0.5913 -4.3828 -4.3828 (0.3698) (0.3671) (0.3738) (0.3678) (3.5634) (3.5634) Kleibergen-Paap F statistic 20.5316 19.8830 17.8550 17.2479 20.7314 20.0738 17.8694 17.2586 4.9053 4.9053 1.7277 1.7277 Observation 2383 2383 2281 2281 2380 2380 2278 2278 919 919 888 888 Medium and Large Firms Outages (log) -0.1003 -0.2131 -0.5988 -0.6520 -0.1252 -0.1252 (0.7733) (0.7551) (0.5686) (0.5745) (0.6510) (0.6510) Outage Hours (log) -0.0673 -0.1296 -0.4256 -0.4403 -0.1101 -0.1101 (0.5287) (0.4823) (0.3993) (0.3881) (0.6582) (0.6582) Kleibergen-Paap F statistic 36.4173 36.3854 10.0050 13.5108 34.9038 34.6161 12.4945 14.6444 35.0721 35.0721 15.7554 15.7554 Observations 1406 1406 1336 1336 1363 1363 1294 1294 697 697 668 668 Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Notes: Small firms are defined as firms with less than 20 workers; medium-sized firms are those with 20-99 employees; while large are firms with 100 and above employees. Firm Controls include: age of the firm, whether the firm is foreign or domestic, and the mobile phone coverage rate in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 81 Table A14: Electricity outages and Firm Performance: Energy Intensive vs Non Energy Intensive Firms Dependent Var: Log Sales Log Sales/Worker Log Value Added/Worker (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) IV Regression Energy Intensive Firms Outages (log) -0.8960 -0.8960 -1.1627 -1.1627 -2.5213∗∗∗ -2.5213∗∗∗ (0.7615) (0.7615) (0.7390) (0.7390) (0.8917) (0.8917) Outage Hours (log) -0.7470 -0.7470 -1.0285 -1.0285 -2.8397 -2.8397 (0.7204) (0.7204) (0.7321) (0.7321) (1.7434) (1.7434) Kleibergen-Paap F statistic 34.5902 34.5902 10.4918 10.4918 35.2353 35.2353 10.2954 10.2954 22.9982 22.9982 5.1918 5.1918 Observation 1372 1372 1297 1297 1362 1362 1288 1288 1132 1132 1082 1082 Non Intensive Firms Outages (log) -0.9944 -1.0381 -1.1371 -1.1692 9.1368 9.1368 (0.6943) (0.7108) (0.7621) (0.7745) (30.3763) (30.3763) Outage Hours (log) -0.4904 -0.5030 -0.5776 -0.5866 -3.8913 -3.8913 (0.3701) (0.3712) (0.4194) (0.4199) (9.8933) (9.8933) Kleibergen-Paap F statistic 18.2784 17.7852 12.1643 12.6947 20.8096 20.3414 13.4956 13.8865 0.0892 0.0892 0.1681 0.1681 Observations 2417 2417 2320 2320 2381 2381 2284 2284 484 484 474 474 Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Notes: Firm Controls include: age of the firm, whether the firm is foreign or domestic, and the mobile phone coverage rate in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 82 Table A15: Effects of Electricity outages on Labor Demand: Energy Intensive vs Non Energy Intensive Firms First Stage Dep. Var: Log # of Full Time Workers Log # of Temp. Workers Log Labor Cost Log Labor Cost per Worker (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) High Energy Intensive Firms Outages (log) 0.3436 0.3436 -0.8156∗∗∗ -0.8156∗∗∗ -0.9421 -0.9421 -0.9115 -0.9115 (0.4425) (0.4425) (0.2957) (0.2957) (0.6967) (0.6967) (0.5553) (0.5553) Outage Hours (log) 0.3926 0.3926 -0.8946∗∗ -0.8946∗∗ -0.8142 -0.8142 -0.7927 -0.7927 (0.4617) (0.4617) (0.4258) (0.4258) (0.7134) (0.7134) (0.5199) (0.5199) Kleibergen-Paap F statistic 32.3987 32.3987 8.8627 8.8627 32.3987 32.3987 8.8627 8.8627 36.0380 36.0380 8.2213 8.2213 37.1387 37.1387 8.0901 8.0901 Observations 1593 1593 1508 1508 1593 1593 1508 1508 1383 1383 1317 1317 1381 1381 1315 1315 Non Energy Intensive Firms Outages (log) -0.2330 -0.2250 -0.3677 -0.3557 -1.0516 -1.0488 -1.1481 -1.1557 (0.2369) (0.2362) (0.2510) (0.2532) (0.7752) (0.7775) (0.8722) (0.8781) Outage Hours (log) -0.1070 -0.1025 -0.1808 -0.1740 -0.5547 -0.5438 -0.6072 -0.6011 (0.1105) (0.1095) (0.1129) (0.1133) (0.4138) (0.4058) (0.4767) (0.4689) Kleibergen-Paap F statistic 15.0938 14.8244 15.3753 15.8048 15.0938 14.8244 15.3753 15.8048 17.8224 17.4242 11.1412 11.6442 17.7279 17.3321 11.1482 11.6461 Observations 2688 2688 2568 2568 2688 2688 2568 2568 2426 2426 2333 2333 2422 2422 2329 2329 Firm Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Climate Ctrls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry FE Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Industry×Year FE No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes Country FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year FE Yes No Yes No Yes No Yes No Yes No Yes No Yes No Yes No Notes: Firm Controls include: age of the firm, whether the firm is foreign or domestic, and the mobile phone coverage in the city of the firm. Climate controls include the log of total precipitation and mean annual temperature. Standard errors clustered at the city level included in parenthesis. ∗ Significant at 10 percent level ∗∗ Significant at 5 percent level ∗∗∗ Significant at 1 percent level 83