Policy Research Working Paper 10510 Power Constraints and Firm-Level Total Factor Productivity in Developing Countries Ablam Estel Apeti Alpha Ly Development Economics Development Research Group June 2023 Policy Research Working Paper 10510 Abstract This paper analyzes the effects of power outages and con- the severity of self-reported power constraints or obstacles straints on manufacturing firms’ revenue-based total factor by firms and the magnitude of revenue-based total factor productivity in developing countries. The empirical analysis productivity loss. The results also indicate that the effect is based on the World Bank Enterprise Surveys dataset for of power outages on firm-level revenue-based total factor 84 countries over 2006–2019. The paper starts by show- productivity could be influenced by the stage of economic ing statistically that firms facing power outages differ and development (low-income countries, lower-middle-income operate in very different environments compared to firms countries, upper-middle-income countries), and the abil- not facing power outages, underlining a potential nonran- ity of firms to engage in research and development and dom issue of the treatment variable. The matching-based purchase backup generators. These findings suggest that approach (entropy balancing) is designed to contain this to ensure economic development, the government should type of bias. It shows that power outages negatively and provide a stable power supply that can mitigate the negative significantly affect firm-level revenue-based total factor pro- shocks faced by manufacturing firms and enhance their ductivity, with a 9 percent lower unconditional average productivity and competitiveness, allowing them to drive productivity for exposed firms compared to nonexposed economic growth. firms. Moreover, the estimates suggest a connection between This paper is a product of the Development Research Group, Development Economics. 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/prwp. The authors may be contacted at aly6@worldbank.org or aapeti@ifc.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 Power Constraints and Firm-Level Total Factor Productivity in Developing Countries* Ablam Estel Apeti† and Alpha Ly‡ JEL Classification: L94, D24, O10 Keywords: Power constraints, Productivity, Developing countries * We thank Raja Chakir (INRAE), Jean-Louis Combes (LEO-UCA), Anna Creti (Paris Dauphine), Carolyn Fischer (The World Bank - DECSI), Megan Elizabeth Lang (The World Bank - DECSI), and Alexandru Minea (LEO-UCA) for their helpful comments and suggestions. We are grateful to the WBG Africa Fellowship Program organizing team. † Ablam Estel Apeti: University of G¨ ottingen, Germany, Universit´ ottingen, G¨ e d’Orl´ e Clermont Auvergne, Universit´ eans, LEO, The World Bank Group, E-mail: ablam estel.apeti@uca.fr. ‡ Alpha Ly (Corresponding Author): Paris Dauphine University, Climate Economics Chair, Chaire EIEA (Mines Paris - PSL & UM6P), Energy and Prosperity Chair, The World Bank - DECSI, E-mail: aly6@worldbank.org / alpha.ly@dauphine.eu 1 Introduction Low levels of infrastructure development and poor quality services within a country can increase production costs for domestic firms and divert their technological choices to sub-optimal solutions. This reduces their level of competitiveness compared to foreign competitors in general. Furthermore, the economic literature suggests that better power infrastructure significantly stimulates economic growth and improves a range of development outcomes. However, in developing countries, firms generally have difficulty connecting to the power grid or, when they are connected, they face frequent scheduled or unscheduled power outages (Alam, 2013). Voltage fluctuations and the frequency of power outages therefore lead to material losses and have a negative effect on manufacturing costs and production. As drivers of economic growth, a major part of the firms in these developing countries cite power as the major obstacle or one of the main constraints to their activities (Asiedu et al., 2021). Indeed, power is the second most important constraint after access to finance. Two regions are particularly affected, namely South Asia and Sub-Saharan Africa. In fact, about 45.9% of firms in Sub-Saharan Africa and 41.2% of firms in South Asia report that power is a major or severe constraint to their operations. The issue of power thus appears to be of greater concern to these firms than many other issues such as corruption and transport in most of these regions. Given the major role of manufacturing firms in developing countries and the gradual improvement of firm-level data availability and quality, a recent literature is emerging on the empirical assessment of the effect of power constraints on firms’ performance. However, this literature on power constraints and firms’ productivity is still scarce and the results remain rather inconsistent when compared to each other. For example, some studies find a statistically significant negative effect of these constraints on the performance of firms (Hardy and McCasland, 2021; Abeberese et al., 2021), other studies find a statistically significant but weak effect (Grainger and Zhang, 2017) and still others do not find a statistically significant relationship between these power obstacles and the productivity of firms (Scott et al., 2014). This lack of consistency in the previous results from the literature could reflect the potential limitations of the different empirical approaches adopted in this literature so far. For example, Xiao et al. (2022) consider power outages or constraints as a completely exogenous explanatory variable. Meanwhile, authors like Fisher-Vanden et al. (2015); Allcott et al. (2016); Cole et al. (2018) or Elliott et al. (2021) consider this variable as potentially endogenous and propose an instrumental variables technique based on variations in the power supply from hydroelectric generation as an instrument. This paper contributes to the empirical literature on the potential effects of power outages or constraints on firm-level TFP by proposing a new matching-based approach (Entropy Balancing) and by extending the previous results from the literature to the degree of severity of the power constraints. First, contrary to Xiao et al. (2022) who present the power outages variable as an exogenous energy shock, we show statistically that the outages treatment variable is not random for manufacturing firms in developing countries. Second, in order to compensate for the limits linked to the instrumentation technique by the hydroelectric variable in some analysis of this literature (as in Fisher-Vanden et al., 2015; Allcott et al., 2016; Cole et al., 2018; Elliott et al., 2021) and to properly account for potential endogeneity with respect to exposure to power constraints, but also to address the lack of a balanced panel structure (survey data), we use a matching-based approach (Entropy Balancing). Our analysis is based on the idea that exposure to constraints represents 2 a treatment. The firms exposed to the constraints constitute the treatment group, the unexposed firms constitute a potential control group. Third, we establish a strong link between the severity of self-reported power constraints or obstacles by firms (minor, moderate, major, severe, and biggest) and the magnitude of productivity loss for firms. In other words, the greater the level of power constraints self-reported by a firm, the greater the effect of these constraints on its productivity, and vice versa. Broadly, our approach shows that the overall effect of power outages on revenue-based total factor productivity (TFPR) is negative and statistically significant. Additionally, we show that other mechanisms such as the acquisition of back-up generators, or investment in R&D allow firms to fight against the severe constraints encountered in the power sector. We also show that constraints in the power sector affect firm revenue-based total factor productivity mainly through the channel of non-optimal reorganization of firm operations (reduced capacity utilization) and through the channel of production losses due to power outages. The rest of the paper is as follows: the literature review in section 2, the theoretical framework in section 3, data and model in section 4, empirical results in section 5 and section 6 concludes. 2 Related literature Some analyses have focused on the effect of inputs market constraints on productivity (Prescott, 2002; Hsieh and Klenow, 2009; Jones, 2011). These analyses suggest that input market constraints prevent the achievement of an efficient allocation of resources, thus reducing total factor productivity. In the same vein, Brandt et al. (2013) find that TFP losses due to input market constraints are still high in many countries. More specifically in our field of research, energy market constraints are also assessed. Indeed, Shi and Sun (2017) find that energy price instability negatively affects output growth in the short and long run. As for Bernanke (1983), he shows that uncertainty about energy prices can induce firms to postpone their investment decisions, thus leading to a decline in overall output. Similarly, Elder and Serletis (2009) suggest that oil price uncertainty may tend to reinforce the negative response of production to oil shocks. Finally, Cheng et al. (2019) show that an increase in oil price volatility reduces real GDP and investment, while a decrease stimulates the economy. With respect to power sector constraints, Cole et al. (2018) find that power outages have a negative and significant effect on firm sales in 14 African countries. The effect found is larger when endogeneity is taken into account.1 They also find that power outages affect firm profits and total factor productivity. Furthermore, also using a hydro-instrumental variable strategy, Elliott et al. (2021) show that Vietnamese firms with less reliable power have lower productivity in 2005 and 2015. They conclude that reducing the length of power outages by 1% would have increased overall revenues by 4.66 billion USD. In the same line of instrumental variables, Allcott et al. (2016) also use changes in power supply from hydroelectric generation as an instrument to estimate the effect of power outages on the Indian manufacturing sector. Finally, a similar approach is taken by Fisher-Vanden et al. (2015) who examine the effects of power outages on firm performance in China. Meanwhile, the economic literature remains relatively well-supplied on the other common determinants of firm-level performance. Among these determinants, we can mention international development aid, fi- nancial inclusion, bank concentration, financial innovation, inflation, or taxation. International aid is a key 1 Use of a hydro-instrumental strategy. 3 factor in improving the performance of firms by alleviating infrastructure and financing constraints in de- veloping countries. In this sense, Chauvet and Ehrhart (2018) find a positive effect of foreign aid on the growth of firms’ sales. Furthermore, Chauvet and Jacolin (2017) find that financial inclusion, i.e. the dis- tribution of financial services in firms, has a positive effect on firm growth. This positive effect is amplified when there is greater competition between banks. In the same vein, Lee et al. (2020) show that financial inclusion helps firms to increase their sales growth. On the other hand, they paradoxically find that finan- cial innovation has a negative effect on the growth rate of firms’ sales. In their paper, Bambe et al. (2022) show that inflation targeting increases the growth and productivity of firms in targeted countries compared to non-targeted countries. Indeed, inflation targeting improves the performance of developing countries by reducing the level and volatility of inflation (Lin and Ye, 2009). Finally, Chauvet and Ferry (2021) show that tax revenues boost the performance of firms through the financing of essential infrastructure for business development. In this paper, we mainly contribute to this literature on the effects of power outages or constraints by proposing a new matching-based approach (Entropy Balancing) and by extending the previous results from the literature to the degree of severity of the power constraints. 3 Theoretical framework The main purpose of this theoretical framework is twofold. First, we highlight the difficulty of directly drawing conclusions about how energy shocks would affect firm-level total factor productivity from a the- oretical framework. The second main objective is to establish theoretically that productivity loss increases with the severity of energy constraints facing manufacturing firms. Although some of the studies in this literature point out that economic growth, for instance, decreases in the event of an energy shock (Cheng et al., 2019; Sadorsky, 1999), we show in this theoretical frame- work that this could have rather contrasting effects on the total factor productivity of manufacturing firms depending on the market structure. Since power constraints are an important manifestation of input market distortion, our theoretical framework is closely related to the literature on the relationship between input market distortion and firm-level total factor productivity. This theoretical framework of intermediate goods and output price adjustment caused by an energy shock is based on that of Hsieh and Klenow (2009) and Xiao et al. (2022). We derive the expression for total factor productivity at the firm-level and analyze the effect of energy shocks on firm-level TFP. We assume monopolistic competition with heterogeneous firms that face different degrees of energy constraints. Indeed, there are several reasons why we consider power constraints to be firm-specific rather than sector-specific in this framework. First, the power infrastructure of a country or region can vary considerably in terms of reliability and quality. Some firms may be located in areas with better power infrastructure, while others may suffer frequent constraints like outages due to inadequate infrastructure or maintenance. Second, some firms may have invested in backup generators or alternative power sources to mitigate the impact of power constraints. These firms would be less affected by general disruptions to the power supply than those without such back-up systems. Third, firms in the same sector may have different operational requirements and production processes. Some firms may be highly dependent on a continuous power supply, such as those involved in refrigeration or other electricity-intensive operations. On the other 4 hand, some firms may have more flexible production processes that can adapt to temporary power supply interruptions. Finally, firms may adopt different mechanisms to cope with power constraints. For instance, some firms may adjust their production schedules, switch to off-peak hours or implement energy-saving measures during outages, helping them to minimize the negative impact of power disruptions. In a fully competitive final output market, there is a single final product Y , and it is aggregated from the output Ys of S manufacturing industries (or branches) by the following production function: S S Y = ∏ Ys θs , where ∑ θs = 1 (1) s=1 s=1 where θs is an output elasticity parameter for industry s. Minimizing production costs implies: PsYs = θs PY (2) θs Ps Ps is the price of the output Ys of the industry s, and the price of the final product is P ≡ ∏S s=1 θs . The output Ys of the industry s is the CES aggregate of Ms types of differentiated products: σ Ms σ −1 σ −1 Ys = ∑ Ysi σ (3) i=1 where σ is the elasticity of substitution between the differentiated products within the industry s. The production function of each differentiated product is given by a Cobb-Douglas function of the firm’s TFP and three production inputs, including capital, labor and energy: αs s β 1−αs −βs Ysi = Asi Ksi Lsi Esi (4) where αs denotes the capital share of industry s, βs denotes the labor share of industry s, (1 − αs − βs ) denotes the share of intermediate products or the share of energy of industry s, and Asi denotes the level of TFP of firm i. Note also that this framework allows input shares to differ between different industries but not between firms in the same industry. The profit maximization program for monopolistically competitive firms with respect to Ksi , Lsi , and Esi is as follows: max πsi = PsiYsi − rKsi − wLsi − (1 + τEsi ) PE Esi (5) s.t. Ysi = Asi Ksi αs Lsi βs Esi 1−αs −βs where r represents the price of capital, w represents the labour cost, PE represents the cost of intermediate inputs, i.e. energy. τEsi represents the energy constraints (increase in the cost of energy, power outages leading to additional costs for back-up generators, increase in losses for firms, etc.) facing a representative firm i. The maximization of the profit (5) using the Lagrangian method of optimization combined with the first order conditions yields: 5 σ −1     λ Asi αs Ksi αs −1 Lsi βs Esi 1−αs −βs − r = 0 where λ = Psi (6)   σ   w αs Ksi = × Lsi (7)  r βs  w 1 − αs − βs 1    Esi = × Lsi (8)   PE βs (1 + τEsi ) By substituting the equations (7) and (8) into the equation (6), we obtain the standard condition that the firm’s output price is a fixed markup over its marginal cost (r, w, PE ): 1−αs −βs σ r αs w βs PE (1 + τEsi )1−α s−βs Psi = × × (9) σ −1 αs βs 1 − αs − βs Asi The output price is increasing with the energy constraints τEsi facing the representative firm. Under the assumption that the output price increases with the level of energy constraint facing the representative firm, this could lead to a generalized price increase in the output market if a significant number of firms face energy constraints in the economy. Indeed, many analyses point out that energy shocks generate output prices adjustment by producers in the economy (Barth III and Ramey, 2001; Bodenstein et al., 2011; Choi et al., 2017).2 This is called cost-push inflation or output market distortion. Depending on the strength of the interactions between the different sectors of the economy, a distortion of the output market as a result of energy constraints may lead to distortions of the labor, capital and even energy markets again, thus leading to a dangerous vicious circle. Now, let us recall our definition of the TFP. We opted for the revenue-based TFP (TFPR). Here, the TFP is measured by revenue productivity or revenue-based productivity such that TFPRsi = Asi Psi (as in Hsieh and Klenow, 2009). From equation (9), we can therefore express the TFPRsi of firm i in the sector s as follows: αs βs 1−αs −βs σ r w PE TFPRsi = × × (1 + τEsi )1−αs −βs (10) σ −1 αs βs 1 − αs − βs The analysis of the effect of energy constraints τEsi on the equation (10) revenue-based TFP leads to: αs βs 1−αs −βs ∂ TFPRsi σ r w PE 1 − αs − βs = × × (11) ∂ τEsi σ −1 αs βs 1 − αs − βs (1 + τEsi )αs +βs Based on the value taken by the elasticity of substitution σ in monopolistic competition, the final expression obtained in equation (11) is either negative (when σ < 1), positive (when σ > 1), or indefinite (when σ → 1). Therefore, it is not possible to draw here any conclusion about how energy shocks will affect total factor productivity at the firm-level from a theoretical perspective (H1 ). That means for instance, the analysis of the effect of power constraints on the productivity of firms thus remains an empirical question that we will analyze in the next section 4. Furthermore, as the limτEsi →+∞ ∂ TFPR ∂ τE si converges toward zero, we can also hypothesize that the greater si the energy constraints facing firms (τEis is large), the more the reverse effects of power constraints on TFP are significant (H2 ). For instance, this would mean that the greater the level of power constraints faced by a firm, the greater the effect of these constraints on its productivity, and vice versa. 2 For example, for Choi et al. (2017), a 10% increase in international oil prices raises inflation by 0.4 percentage points on average. 6 The main purpose of our following empirical approach will be to test these two main hypotheses (H1 and H2 ) put forward through this theoretical framework. 4 Empirical strategy and data 4.1 Empirical strategy One of the main challenges in quantifying the effect of power sector constraints on firm performance is that facing or not power constraints might be non-random (Alam, 2013; Cole et al., 2018; Elliott et al., 2021). In- deed, the power constraints faced by firms in a country can be explained by the macroeconomic/institutional phenomena of the country in question (low quality of regulation, lack of financing, political instability, etc.) and/or firm characteristics such as size, maturity, quality of management or even ownership (public or pri- vate, domestic or foreign). These realities can also explain the low level of performance of firms operating in that country. Even within the same country, we can mention targeted public investment in energy infras- tructure near the best performing firms to support their operations, and public investment in infrastructure (roads and railways) that can both improve the reliability of power supply (ease of maintenance of power transmission lines) and the transportation of products for firms. To deal with this identification issue, some authors in this literature have opted for the instrumental variables technique based on variations in the power supply from hydroelectric generation as an instrument (Fisher-Vanden et al., 2015 in China; Allcott et al., 2016 in India; Cole et al., 2018 in 14 selected Sub- Saharan African countries; Elliott et al., 2021 in Vietnam). Meanwhile, we believe that this identification strategy might potentially have three major limitations: First, it is not applicable to a larger sample, as not all countries have an energy mix based mainly on hydropower. As a result, studies using this identification strategy frequently focus on a single country, or on a small number of countries similar in terms of energy mix mainly based on hydro-generation of the electricity, thus reducing the external validity of their conclusions. The second and more fundamental limitation might be that very often, these countries that are largely de- pendent on hydro-generation set up substitute generators or fossil power plants ready to take over following the fluctuations in dam water (climate shocks). This reality could cut the link between hydro-generation (the instrument) and power constraints (treatment). The validity of this type of instrument would therefore be partly questioned. Finally, even when a valid instrument is available, the fact that the exposure to the treatment is non-random could result in poor estimates when using an instrumental variable strategy approach, as pointed out by Ertefaie et al. (2016) and Canan et al. (2017), who stress the relevance of the matching approach in the presence of this specific source of endogeneity. So, to overcome these potential limitations described above and to properly account for potential endo- geneity issue with respect to exposure to power constraints, but also to address the lack of a balanced panel structure (survey data), we propose a matching-based approach (Entropy Balancing). Our analysis is based on the idea that exposure to constraints represents a treatment. Firms exposed to the constraints constitute the treatment group, while firms not exposed constitute a potential control group. The average treatment effect on treated firms (ATT) is defined as follows: 7 τATT = E [y(1) | T = 1] − E [y(0) | T = 1] (12) y(·) is the outcome variable (the TFPR). T indicates whether a unit or firm is exposed to treatment/constraints (T = 1) or not (T = 0). Therefore, E [y(1) | T = 1] is the expected outcome after treatment and E [y(0) | T = 1] is the counterfactual outcome, i.e. the outcome that a unit exposed to treatment would have obtained if it had not experienced to treatment. As the counterfactual outcome is not observable, we need an appropriate proxy to identify the ATT. If the treatment is randomly assigned, then the average outcome of the units not exposed to the treatment, E [y(0) | T = 0], is an appropriate proxy. However, as we saw earlier, exposure to the constraints and hence selection into the treatment could be endogenous due to the potential confounding factors we mentioned earlier. The idea of matching-based estimators is to mimic randomization with respect to treatment assignment. The unobserved counterfactual outcome is imputed by matching treated units with untreated units that are as similar as possible with respect to all observable characteristics that: (i) are associated with selection into treatment (i.e. the probability of being exposed to constraints), and (ii) influence the outcome of interest. The realizations of the productivity gap measure for these matches are then used as an empirical proxy for the unobserved counterfactual. Formally, the matching-based ATT estimate is defined as follows: τATT (x) = E [y(1) | T = 1, X = x] − E [y(0) | T = 0, X = x] (13) where x is a vector of relevant observable characteristics (see the description of the firm-level and country- level control variables in subsection 4.2), E [y(1) | T = 1, X = x] is the expected outcome for the units that received the treatment, and E [y(0) | T = 0, X = x] is the expected outcome for the best matches of the treated units. Entropy balancing estimates the causal effect under the unconfoundedness assumption or conditional independence assumption (CIA). The CIA implies that the selection into the treatment group is only conditional to a set of observed covariates. Specifically, it assumes that conditional on the observed covariates (after controlling-for the covariates), the treatment assignment is independent of the potential outcomes.3 In this study, as Neuenkirch and Neumeier (2016) in the analysis of the effect of US sanctions on the poverty gap in the target countries and Apeti (2023a) in the analysis of the effects of mobile money on household consumption volatility, we use Entropy Balancing to select matches for units exposed to the treatment and to estimate the ATT.4 Entropy Balancing is a method proposed by Hainmueller (2012). This method is implemented in two steps. First, weights are calculated and assigned to the units not subject to treatment. These weights are chosen to satisfy prespecified equilibrium constraints involving sample moments of the observable features while at the same time remaining as close as possible to the uniform base weights. In our analysis, the equilibrium constraints require equal covariate means between the treatment and control groups, which ensures that the control group contains, on average, non-treatment units that are as similar as possible to the treatment units. Second, the weights obtained in the first step are used in a regression analysis with the treatment indicator as an explanatory variable. This yields an estimate of the ATT, i.e. the conditional difference in the means of the outcome variable between the treatment group and 3 In other words, CIA implies that after conditioning on the observed covariates, there are no unobserved confounding factors that influence both the treatment assignment and the potential outcomes. We empirically test the CIA in our subsection 5.2. 4 See also Apeti and Edoh (2023); Apeti (2023b); Apeti et al. (2023). 8 the control group.5 Broadly, the idea of Entropy Balancing here is to compare the productivity gap of firms exposed to power constraints with that of unexposed firms that are as similar as possible to the exposed firms. The average difference in productivity between the exposed firms and the ”closest” unexposed firms must then be due to the treatment, i.e. the exposure to the power constraints. In this sense, the empirical approach mimics a randomized experiment by balancing the treatment and control groups on the basis of observable characteristics. By combining matching and regression analysis, Entropy Balancing has some advantages over other treatment effect estimators. A particularly important advantage over regression-based approaches (includ- ing DiD estimation) as well as propensity score-based matching methods is that Entropy Balancing is non- parametric in the sense that no empirical model for the outcome variable or selection into treatment needs to be specified. Moreover, unlike regression-based analyses, there is no multicollinearity, as the reweighting scheme orthogonalizes the covariates to the treatment variable. Furthermore, unlike other matching meth- ods, Entropy Balancing ensures a high balance of covariates between treatment and control groups, even in small samples. Then, by combining a reweighting scheme with regression analysis, Entropy Balancing allows us to control for both country fixed effects and year fixed effects in the second step of the matching approach, i.e. the regression analysis.6 The inclusion of country fixed effects is particularly useful to account for the potential unobserved heterogeneity between firms from different countries and to control for time-invariant country-specific conditions that could lead to differences in the productivity gap between firms. Also, knowing that productivity varies with firm and economic characteristics (Syverson, 2011), we include a large set of control variables at the firm and country level. Finally, as recalled by Chauvet and Ehrhart (2018), the statistical bias resulting from the attempt to measure the effect of macro variables on micro-units was underlined by Moulton (1990).7 Therefore, as in Chauvet and Ehrhart (2018) and Bambe et al. (2022), the standard errors are clustered at the country level (as we have country level control variables such as economic growth, bank concentration, inflation, etc.).8 4.2 Data and variables In this analysis, we consider 31,406 manufacturing firms in 84 developing countries from 2006 to 2019, 30 of which are in Sub-Saharan Africa (AFR), 8 in East Asia and Pacific (EAP), 14 in Europe and Central Asia (ECA), 20 in Latin America and the Caribbean (LAC), 6 in the Middle East and North Africa (MNA), and 6 in South Asia (SAR). 4.2.1 Firm-level data We mainly use the World Bank Enterprise Survey (WBES) data in this analysis. The strata of the enterprise surveys are firm size, industry and geographical region within a country. The firm size levels are 5-19 (small), 20-99 (medium), and 100+ employees (large). It is important to note, however, that these surveys 5 In the regression step, we additionally control for the covariates used in the first step. This is equivalent to including control variables in a randomized experiment to increase the efficiency of the estimation. 6 We also include region and income group fixed effects in our econometric specification. 7 Random disturbances in the correlated regression within the groupings that are used to merge the aggregate data with the micro data can lead to a downward bias in the ordinary least squares standard errors (Moulton, 1990). 8 Clustering the standard errors at the industry level also yields similar results as the country level clustering (see robustness check in the subsection 5.3). 9 are limited to formal firms with five or more employees and with over 1% private ownership or participation. In this analysis, we start with the year 2006 because most surveys from this year onwards use stratified sampling and contain weights based on this information. Earlier surveys may not contain any information on weights. All monetary variables are adjusted to 2009 levels using a World Bank GDP deflator and then transformed to US dollars using the International Monetary Fund’s purchasing power parity (PPP) index. Our treatment variables (dummy variables taking 0 or 1) characterize the different levels of power con- straints that firms report being exposed to. One advantage of this measure is that it reflects firms’ perceptions of the extent to which power constrains their operations. This is important because the firm’s perception is one of the most important factors influencing their operational and investment decisions (Asiedu and Freeman, 2009). To approximate the level of performance of each firm, we opt for total factor productivity (TFP), the part of output that is not explained by the quantity of inputs used.9 The productivity of firms, i.e. the ability to generate greater output with fewer inputs, is one of the key elements of economic growth. As a reminder, productivity is estimated from a Cobb-Douglas VA (value added) function of the following form: α α VAi = Ai Ki k Li l (14) where value added at the firm-level VAi is a function of the inputs of capital (Ki ) and labour (Li ).10 The efficiency of firms’ production is measured by the term Ai which is the part of output that cannot be directly attributed to inputs used. Equation (14) could be rewritten as: log(VAi ) = α0 + αk log(Ki ) + αl log(Li ) + εi (15) log(TFPi ) is estimated as a sum of the constant and the residual, i.e., log(TFPi )=α ˆi .11 Therefore, the ˆ0 + ε TFP is the part of output that is not explained by the quantity of inputs used. In equation (14), TFP Ai is estimated separately for each industry. This avoids the assumption of a common production technology (i.e. αk and αl are the same within the sample). In addition, wherever possible, the elasticities of output with respect to capital and labor (i.e. αk and αl ) can vary according to the income level group of the corresponding economy. Finally, as in Halvorsen et al. (1980), the country and year effects are controlled via dummy variables for each country and year. The table (1) shows an average log(TFP) of 2.46 for our sample of firms. However, we have a large heterogeneity between our firms as it varies from -3.23 for the least productive firms to 8.83 for the best performing firms. At the firm-level, we control for the age of the firm, the size of the firm, whether or not the firm has a website, and most importantly, the ownership of the firm.12 Indeed, over time, firms tend to find ways to solve or mitigate the problem of power constraints. In addition, older firms are found to be more productive (Majumdar, 1997). Moreover, a positive relationship between firm size and TFP is found in the manufac- turing sector in general (Leung et al., 2008; Tovar et al., 2011). Also, we use the possession of a website by a firm as a good signal of the quality of the firm’s marketing, it can influence its productivity. Finally, 9 World Bank Group, Enterprise Analysis Unit. 2017. “firm-level Productivity Estimates”. 10VA is represented by the difference between the establishment’s total annual sales and the total annual cost of inputs, K is represented by the replacement value of machinery, vehicles and equipment; L is represented by the total annual labor cost. 11 Where log(A )=log(TFP ) i i 12 We include both age and size because, contrary to popular belief, St-Pierre et al. (2010) have shown that firm size and age are not substitutes in an exploratory study of 288 Quebec manufacturing firms. 10 Table 1: Summary statistics on our main variables Mean SD Min Max N Panel A: Firm-level controls Firm’s longevity (Years) 29.72 16.77 3.00 351.00 31406 Firm size (Small=1, Medium=2 or Large=3) 1.81 0.77 1.00 3.00 31406 Own website (No=0, Yes=1) 0.42 0.49 0.00 1.00 31406 Foreign private participation (%) 7.75 24.93 0.00 100.00 31406 Panel B: Country-level controls Regulatory Quality, Percentile Rank (0-100) 42.67 18.55 5.21 91.75 31406 Financial development index 0.28 0.14 0.03 0.70 31406 Bank concentration (%) 61.49 20.04 22.60 100.00 31406 Net ODA received (% of GNI) 2.42 4.16 0.01 37.37 31406 Inflation, consumer prices (annual %) 7.58 8.58 -1.05 84.86 31406 GDP growth (annual %) 4.34 4.44 -25.91 14.01 31406 GDP per capita (constant 2015 USD) 3976.76 3208.01 302.07 14086.02 31406 Panel C: Outages related variables Outages frequency (occurrences per month) 7.31 8.69 0.00 31.00 16821 Outages intensity (hours per occurrence) 4.45 6.98 0.00 96.00 16416 Panel D: Performance related variables Log revenue-based TFP (TFPR) 2.46 1.50 -3.23 8.83 31406 Cost of inputs per unit of sales (2009 USD) 0.40 0.22 0.00 0.99 31406 Sales per labor cost (2009 USD) 11.71 21.32 0.11 646.50 31406 Capacity Utilization (%) 73.48 21.12 0.00 100.00 30714 Losses due to Outages (% of annual Sales) 8.12 11.85 0.00 100.00 12813 Asiedu et al. (2021) find that the probability of facing power constraints is lower for state-owned firms with majority ownership. However, a high level of private participation in a firm could reveal significant levels of firm attractiveness or performance. Further details on the firm-level variables can be found in the Appendix (Table 16). 4.2.2 Country-level data At the country level, we control the quality of regulation, financial development, bank concentration, foreign aid, inflation, economic growth, and the level of wealth. Indeed, Agostino et al. (2020) establish strong evidence that better local institutions (the rule of law and government efficiency) help small and medium- sized firms become more productive in Europe over the period 2010-2014. We also have evidence of a positive effect of foreign aid on firm sales growth (Chauvet and Ehrhart, 2018). Furthermore, financial development and greater competition between banks (strong bank concentration) favor the performance of firms ( Chauvet and Jacolin, 2017; Lee et al., 2020). Finally, Bambe et al. (2022) show that inflation targeting increases the growth and productivity of firms in targeted countries compared to non-targeted countries. Indeed, inflation targeting improves the performance of developing countries by reducing the level and volatility of inflation (Lin and Ye, 2009). Further details on the country-level variables can be found in Appendix Table (16). 11 5 Empirical results 5.1 Descriptive statistics Table (2) presents a summary of our treatment variables (power constraints) in this study. With respect to power outages, 16,816 firms out of 30,403 firms faced power outages, or about 53.6% of the firms in our sample. In addition, for 19,551 firms, power is an obstacle, representing 62.3% of firms. Of these firms, 17.6% faced minor obstacles in the power sector, 15.1% faced moderate obstacles, 16.2% faced major obstacles, 13.2% faced severe obstacles, and 11.7% considered power to be the major challenge in their operations. Finally, while outages are one of the main manifestations of constraints in the power sector in these countries, there are a number of other constraints (high cost of power, voltage fluctuations, connection problems, etc.) that are not readily observable through the survey data. Indeed, we can see here that 32.7% of the firms that faced obstacles did not experience any outages. Similarly, 39.1% of firms that experienced minor obstacles, 28.3% of firms that experienced moderate obstacles, 28.9% of firms that experienced major obstacles, 33.8% of firms that experienced severe obstacles, and 27.2% of firms that reported power as their greatest challenge. Table 2: Summary statistics on the treatment variables s es es e tag tag tag s ges s e e Ou Ou Ou tag tag l l l ta ta ta ta Ou Ou Ou No No No To To To Count Row percentages Column percentages Power is an obstacle No 8,198 3,654 11,852 69.2% 30.8% 100.0% 56.2% 21.7% 37.7% Yes 6,388 13,163 19,551 32.7% 67.3% 100.0% 43.8% 78.3% 62.3% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Power as minor obstacle No 12,420 13,446 25,865 48.0% 52.0% 100.0% 85.1% 80.0% 82.4% Yes 2,167 3,371 5,538 39.1% 60.9% 100.0% 14.9% 20.0% 17.6% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Power as moderate obstacle No 13,242 13,406 26,648 49.7% 50.3% 100.0% 90.8% 79.7% 84.9% Yes 1,345 3,410 4,755 28.3% 71.7% 100.0% 9.2% 20.3% 15.1% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Power as major obstacle No 13,114 13,191 26,305 49.9% 50.1% 100.0% 89.9% 78.4% 83.8% Yes 1,472 3,625 5,098 28.9% 71.1% 100.0% 10.1% 21.6% 16.2% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Power as severe obstacle No 13,182 14,060 27,242 48.4% 51.6% 100.0% 90.4% 83.6% 86.8% Yes 1,404 2,756 4,161 33.8% 66.2% 100.0% 9.6% 16.4% 13.2% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Power as biggest obstacle No 13,097 13,835 26,932 48.6% 51.4% 100.0% 93.1% 84.2% 88.3% Yes 968 2,588 3,556 27.2% 72.8% 100.0% 6.9% 15.8% 11.7% Total 14,065 16,423 30,488 46.1% 53.9% 100.0% 100.0% 100.0% 100.0% N 11,808 18,680 30,488 Table (3) shows that 52.9% of our firms are small firms employing fewer than 20 people, and 35.3% are intermediate firms employing between 20 and 99 people. Finally, large firms (more than 100 employees) represent 11.8% of firms. For each category of these firms, the frequency of experiencing outages exceeds 50 percent. On the other hand, only 19.1% of firms invest in R&D, and of these, a significant 62.3% are roughly firms that suffer from power constraints. Similarly, of the 21.5% of firms that own back-up generators, 77.2% are actually firms that face power constraints. The next figures present some stylized facts. In Figure (1), the white lines in the middle of the boxes 12 Table 3: Summary statistics on our main categorical variables es s es ge tag tag uta es es es Ou Ou tag tag tag O tal tal tal Ou Ou Ou No No No To To To Count Row percentages Column percentages Firm size Small(<20) 7,974 8,639 16,613 48.0% 52.0% 100.0% 54.7% 51.4% 52.9% Medium(20-99) 4,984 6,087 11,071 45.0% 55.0% 100.0% 34.2% 36.2% 35.3% Large(100 And Over) 1,629 2,091 3,719 43.8% 56.2% 100.0% 11.2% 12.4% 11.8% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% R&D No 8,726 9,492 18,219 47.9% 52.1% 100.0% 84.4% 78.0% 80.9% Yes 1,617 2,672 4,289 37.7% 62.3% 100.0% 15.6% 22.0% 19.1% Total 10,344 12,164 22,508 46.0% 54.0% 100.0% 100.0% 100.0% 100.0% Generator No 13,016 11,596 24,612 52.9% 47.1% 100.0% 89.5% 69.0% 78.5% Yes 1,533 5,202 6,735 22.8% 77.2% 100.0% 10.5% 31.0% 21.5% Total 14,549 16,798 31,347 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Own website No 9,327 10,983 20,309 45.9% 54.1% 100.0% 63.9% 65.3% 64.7% Yes 5,260 5,834 11,094 47.4% 52.6% 100.0% 36.1% 34.7% 35.3% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% WB income group Low Income 281 1,386 1,666 16.8% 83.2% 100.0% 1.9% 8.2% 5.3% Lower Middle Income 7,634 10,411 18,045 42.3% 57.7% 100.0% 52.3% 61.9% 57.5% Upper Middle Income 6,672 5,020 11,692 57.1% 42.9% 100.0% 45.7% 29.9% 37.2% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% Region AFR 2,542 3,075 5,616 45.3% 54.7% 100.0% 17.4% 18.3% 17.9% EAP 3,790 4,315 8,105 46.8% 53.2% 100.0% 26.0% 25.7% 25.8% ECA 1,605 967 2,573 62.4% 37.6% 100.0% 11.0% 5.8% 8.2% LAC 2,993 2,747 5,740 52.1% 47.9% 100.0% 20.5% 16.3% 18.3% MNA 2,076 2,487 4,563 45.5% 54.5% 100.0% 14.2% 14.8% 14.5% SAR 1,581 3,225 4,806 32.9% 67.1% 100.0% 10.8% 19.2% 15.3% Total 14,587 16,816 31,403 46.4% 53.6% 100.0% 100.0% 100.0% 100.0% N 12,317 19,086 31,403 indicate the respective medians. After a quasi-constant evolution within minor and moderate obstacles, we observe a drop in median total factor productivity as soon as the constraints become major for the firms (box plots 1). It can be seen that this deterioration in the power sector lowers the median total factor productivity of firms. Similarly, the box plots 2 shows a decrease in median firms’ capacity utilization with the level of power constraints. We can see that this deterioration in the power sector reduces the median capacity utilization. The box plots 3 shows that the median losses incurred by firms increases with the level of constraints encountered in the power sector. The deterioration in the power sector increases median losses. Figure (2) shows the effects of constraints on losses with a breakdown by geographical region. We can see that the negative effects of power constraints on losses by region are reinforced when the constraints become major or severe, especially for the Sub-Saharan Africa, MENA and South Asia regions. Figure (3) highlights the combined effect of outages and other constraints. When the outages are not associated with the highest level of constraints, we can see the decrease in productivity only for two regions (EAP and LAC). However, when the outages are associated with the highest level of constraints, then we find the decrease in productivity in four regions (AFR, EAP, ECA, SAR). Moreover, intra-regional volatility in total factor productivity becomes more important between firms within the same region. Table (4) shows the sample means of all matching covariates, divided into two groups: observations of firms facing power constraints (outages) or the treatment group (column 1) and observations of firms not facing power constraints or the potential control group (column 2).13 The last column shows the standard- 13 As Asiedu et al. (2021), we log some of our macroeconomic variables to mitigate the effect of outliers. 13 100 8 6 80 Total Factor Productivity Capacity Utilization (%) 4 60 2 40 0 -2 20 No obstacle Minor obstacle Moderate obstacle Major obstacle Very severe obstacle No obstacle Minor obstacle Moderate obstacle Major obstacle Very severe obstacle excludes outside values excludes outside values 40 Losses due to Outages (% of annual Sales) 10 20 0 30 No obstacle Minor obstacle Moderate obstacle Major obstacle Very severe obstacle excludes outside values Figure 1: Degree of power constraints, declining total factor productivity and capacity utilization in %, and increasing losses in % of total sales No obstacle Minor obstacle Moderate obstacle 80 60 Losses due to Outages (% of annual Sales) 40 20 0 AFR EAP ECA LAC MNA SAR AFR EAP ECA LAC MNA SAR AFR EAP ECA LAC MNA SAR Major obstacle Very severe obstacle 80 60 40 20 0 AFR EAP ECA LAC MNA SAR AFR EAP ECA LAC MNA SAR excludes outside values Figure 2: Breakdown losses by degree of power constraints and by region 14 2.7 2.7 2.6 2.6 Total factor productivity Total factor productivity 2.5 2.5 2.4 2.4 2.3 2.2 2.3 No Yes No Yes Outages but Power is not the biggest obstacle Outages and Power as the biggest obstacle AFR EAP ECA LAC MNA SAR AFR EAP ECA LAC MNA SAR Figure 3: The combined effects of outages and other power constraints ized differences in means between these two groups with the corresponding level of significance in each case. Table 4: Descriptive Statistics (T=Power Outages) [1] [2] [1] − [2] Outages No Outages Diff. Firm size 1.802 1.831 -0.038*** Firm’s longevity 29.690 29.737 -0.003** Own website 0.386 0.471 -0.173*** Foreign private participation 8.082 7.222 0.034** GDP growth (annual %) 3.321 4.995 -0.511*** Inflation (annual %) 7.726 7.480 0.043*** log(Bank concentration in %) 4.047 4.084 -0.108*** log(FD index) -1.496 -1.243 -0.436*** log(GDP per capita) 7.771 8.231 -0.542*** log(Net ODA in % of GNI) -0.185 -0.883 0.392*** log(Regulatory Quality) 3.601 3.784 -0.370*** Observations 19086 12317 Notes: This Table shows the sample means of all matching covariates, divided into two groups: observations of firms facing outages or the treatment group (column 1) and observations of firms not facing outages or the potential control group (column 2). The last column shows the standardized differences in means between these two groups with the corresponding level of significance in each case. The analysis of the results for all the relevant observable characteristics reveals that firms facing outages differ drastically compared to firms not facing outages. The analysis of the results for all relevant observable characteristics reveals that firms facing power constraints differ drastically and operate in very different environments compared to firms not facing power constraints. Indeed, we find that firms facing power constraints are on average smaller than those not facing power constraints. In the same vein, these firms facing the constraints are on average younger than those not facing the constraints. We also note that firms not facing constraints have on average a better marketing management (approximated by the possession of a website). Finally, we also notice that the firms that face 15 constraints more often in terms of power services are the firms that are mostly owned by foreign private actors. Furthermore, the economic and policy environment in which these power-constrained firms operate is generally worse. Economic growth is lower, inflationary pressures are higher, financial development and banks concentration are lower, and the quality of regulation is weaker. These firms are also found to be located in poorer and less resilient countries and therefore receive more international development assistance and aid. These descriptive results illustrate the importance of selecting an appropriate control group using a matching-based approach before calculating treatment effects, as otherwise the effect of power constraints on firms’ total factor productivity could be miss-estimated. Table (5) compares the sample means of all matching covariates in the treatment group (column 1) and the synthetic control group obtained via Entropy Balancing (column 2). The last column shows the standardized differences in means with the corresponding significance level in each case. The comparison of the average realizations of the observable characteristics of the treatment group with those of the synthetic control group reveals the effectiveness of Entropy Balancing. All covariates are perfectly balanced and no statistically significant differences remain. Furthermore, Figures (4 and 5) display the kernel densities of our covariates for the treatment and control group and show how balancing constraints have affected the reweighted covariate distributions. Therefore, we can say that the control groups in the subsequent empirical analysis are composed of relevant counterfactuals for the sample of firms facing power constraints. Table 5: Covariate balancing (T=Power Outages) [1] [2] [1] − [2] Outages Control Diff. Firm size 1.802 1.802 0.000 Firm’s longevity 29.737 29.736 0.000 Own website 0.386 0.386 -0.000 Foreign private participation 8.082 8.082 0.000 GDP growth (annual %) 4.995 4.994 0.000 Inflation (annual %) 7.480 7.482 -0.000 log(Bank concentration in %) 4.047 4.047 -0.000 log(FD index) -1.496 -1.496 -0.000 log(GDP per capita) 7.771 7.771 -0.000 log(Net ODA in % of GNI) -0.185 -0.185 -0.000 log(Regulatory Quality) 3.601 3.601 -0.000 Weighted observations 19086 19086 Notes: This Table compares the sample means of all matching covariates in the treatment group (column 1) and the synthetic control group obtained via Entropy Balancing (column 2). The last column shows the standardized differences in means with the corresponding significance level in each case. The comparison of the average realizations of the observable characteristics of the treatment group with those of the synthetic control group reveals the effectiveness of Entropy Balancing. All covariates are perfectly balanced and no statistically significant differences remain. 16 Balancing on the 1st order Balancing on the 1st order .04 .111 .03 .11 Kernel density Kernel density .109 .02 .108 .01 treated .107 treated control control 0 0 100 200 300 400 1 1.5 2 2.5 3 Firm's longevity Firm size Balancing on the 1st order Balancing on the 1st order .1 .08 .08 .06 Kernel density Kernel density .06 .04 .04 .02 .02 treated treated control control 0 0 -30 -20 -10 0 10 20 0 20 40 60 80 GDP growth (annual %) Inflation, consumer prices (annual %) Balancing on the 1st order Balancing on the 1st order .1115 .11 .111 .108 Kernel density Kernel density .1105 .106 .11 .104 .102 .1095 treated treated .1 control control .109 3 3.5 4 4.5 -4 -3 -2 -1 0 log(Bank concentration in %) log(Financial development index) Figure 4: Kernel densities of the covariates for the treatment and control group (Set 1) 17 Balancing on the 1st order Balancing on the 1st order .11 .1 .09 .1 Kernel density Kernel density .08 .09 .07 treated treated .08 .06 control control 6 8 10 12 -4 -2 0 2 4 log(GDP per capita) log(Net ODA received in % of GNI) Balancing on the 1st order Balancing on the 1st order .1 .11 .08 Kernel density Kernel density .06 .105 .04 .02 treated treated .1 control control 0 1 2 3 4 5 0 20 40 60 80 100 log(Regulatory Quality) Foreign private participation Balancing on the 1st order .1112 .111 Kernel density .1108 .1106 .1104 treated control .1102 0 .2 .4 .6 .8 1 Own website Figure 5: Kernel densities of the covariates for the treatment and control group (Set 2) 18 5.2 Treatment effects Based on the treatment effects for endogenous treatments described in Wooldridge (2010), we performed an endogeneity test to ensure that the conditional independence assumption (CIA) is respected before run- ning the regressions based on entropy balancing. The results shows that conditional on our covariates, the treatment and outcome unobservables are uncorrelated (H0).14 We can be sure that our estimates below from entropy balancing represent the consistent treatment effects on the treated (ATT). The results of the Table (6) indicate that power constraints characterized by outages negatively affect ex- posed firms, since we observe a negative and significant coefficient associated with power outages dummy variable of 0.102, i.e., 1.1 percentage points or 9% of the unconditional average productivity (column 1). This first result supports those found by a large part of this literature (Cole et al., 2018; Abdisa, 2018; Elliott et al., 2021; Xiao et al., 2022). However, we note that the effects of the constraints are not statistically sig- nificant when the levels of constraints are not severe (columns 2 to 5), even if they are gradually increasing. As soon as the constraints become very important (severe), the effect becomes negative and statistically significant (column 6). In column (7), the adverse effect is greatest when power proves to be the firm’s biggest obstacle (ahead of factors such as access to finance, problems with the tax administration, transport problems or problems related to corruption). The visualization of these coefficients are shown in Figure (6). Table 6: The effect of power outages and constraints on the firm’s revenue-based TFP (1) (2) (3) (4) (5) (6) (7) Power Outages -0.102∗∗∗ (0.032) Power is not an obstacle 0.041 (0.027) Power as minor obstacle 0.040 (0.026) Power as moderate obstacle 0.004 (0.031) Power as major obstacle -0.032 (0.028) Power as severe obstacle -0.044∗ (0.025) Power as biggest obstacle -0.148∗∗∗ (0.025) Observations 31403 31406 31406 31406 31406 31406 30491 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regres- sions (Entropy Balancing). The dependent variable is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-7). In each regression, the treatment variable represents a specific degree of power constraints for firms. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Infla- tion (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. The variation in the number of observations is related to the fact that the treatment variables are derived from three separate survey questions. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 In Table (7), we consider the interaction of the outages treatment variable with each specific level of obstacles encountered by firms in the power sector (the other treatment variables in our analysis). In column (1), when we consider the outages treatment variable and no constraints or obstacles, we have no effect of 14 Test of endogeneity. H0: Treatment and outcome unobservables are uncorrelated. chi2(2) = 0.40 ; Prob > chi2 = 0.8191). We cannot reject H0. 19 Power is not an obstacle Power as minor obstacle Power as moderate obstacle Power as major obstacle No obstacle Power as severe obstacle Minor obstacle Moderate obstacle Major obstacle Power as biggest obstacle Severe obstacle Biggest obstacle -.2 -.1 0 .1 Figure 6: The effect of power constraints on the firm’s revenue-based TFP Notes: Figure shows the coefplots with the smoothed confidence intervals (1, 3, 5, ..., 99) of the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing) from the Table (6). The dependent variable is the logarithm of the firm-level revenue- based total factor productivity for all regressions. In each regression, the treatment variable represents a specific degree of power constraints for firms. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors are clustered at country level. outages on firms’ productivity. This implies that most firms manage to deal with outages when other sources of constraints in the power sector are absent (voltage problems, high cost of service, connection difficulties, etc.). In column (2), the effect becomes negative when we consider the presence of minor constraints. The negative effect becomes progressively stronger when we consider the presence of moderate and major obstacles, until it becomes significant when we consider the presence of severe obstacles for firms. This negative effect is reinforced and becomes very significant when power becomes the greatest obstacle for firms. 5.3 Robustness tests The World Bank study suggests clustering by industry when each industry has at least 500 observations.15 Otherwise, the appropriate clustering is by economies or countries as we have done so far. However, in order to test the sensitivity of our results to this, we have repeated our regressions by clustering by industry (see Table 8) even though we have fewer than 500 observations for some industries in our sample. Although we have some small variations in our standard errors, the results remain almost the same (in quality and magnitude) as our initial results with the clustering by country. We also wanted to test the robustness of our results using the ordinary least squares (OLS) estimator in Table (9) to reassure that our results are not highly influenced or biased by our choice of Entropy Balancing estimator. Obviously, we have some small differences in magnitudes for some of the coefficients due to the downward bias of the OLS estimates in presence of endogeneity (Neuenkirch and Neumeier, 2016), but 15 World Bank Group, Enterprise Analysis Unit. 2017. “firm-level Productivity Estimates”. 20 Table 7: The combined effect of power outages with each level of power constraints on the firm’s revenue- based TFP (1) (2) (3) (4) (5) (6) Outages × No obstacle 0.016 (0.027) Outages × Minor obstacle -0.014 (0.037) Outages × Moderate obstacle -0.035 (0.040) Outages × Major obstacle -0.044 (0.028) Outages × Severe obstacle -0.079∗∗ (0.031) Outages × Biggest obstacle -0.173∗∗∗ (0.026) Observations 31406 31406 31406 31406 31406 31406 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The dependent variable is the logarithm of the firm- level revenue-based total factor productivity for all regressions (1-6). In each regression, the treatment variable represents a specific degree of power constraints for firms and outages dummy variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own web- site, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Coun- try FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Table 8: Robustness tests – The effect of power outages and constraints on the firm’s revenue-based TFP by clustering standards errors at industry level (1) (2) (3) (4) (5) (6) (7) Power Outages -0.102∗∗∗ (0.018) Power is not an obstacle 0.041 (0.025) Power as minor obstacle 0.040 (0.029) Power as moderate obstacle 0.004 (0.025) Power as major obstacle -0.032 (0.022) Power as severe obstacle -0.044∗ (0.022) Power as biggest obstacle -0.148∗∗∗ (0.042) Observations 31403 31406 31406 31406 31406 31406 30491 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regres- sions (Entropy Balancing). The dependent variable is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-7). In each regression, the treatment variable represents a specific degree of power constraints for firms. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Infla- tion (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at industry level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 21 they remain qualitatively unchanged. As Entropy Balancing, the OLS estimator also allows us to see the adverse and progressive effect of power sector constraints on firm productivity in developing countries. Table 9: Robustness tests – The effect of power outages and constraints on the firm’s revenue-based TFP by using least squares regressions (without entropy balancing re-weighting scheme) (1) (2) (3) (4) (5) (6) (7) Power Outages -0.083∗∗∗ (0.028) Power is not an obstacle 0.041 (0.026) Power as minor obstacle 0.031 (0.022) Power as moderate obstacle -0.000 (0.029) Power as major obstacle -0.028 (0.028) Power as severe obstacle -0.047∗∗ (0.024) Power as biggest obstacle -0.131∗∗∗ (0.027) Observations 31403 31406 31406 31406 31406 31406 30491 Notes: Table shows the coefficients obtained by least squares regressions. The dependent variable is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-7). In each regres- sion, the treatment variable represents a specific degree of power constraints for firms. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level con- trols variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 In table (10), we opt for alternative measures of firm productivity. Although our main variable (TFPR) is suitable to approximate the level of performance of manufacturing firms in developing countries, it also faces some criticism in the literature. Indeed, the estimation of the TFPR can potentially be problematic for certain reasons such as selection, simultaneity or problems related to the use of monetary measures (as opposed to physical measures) of production and inputs (Levinsohn and Petrin, 2003, Foster et al., 2008, Hsieh and Klenow, 2009). To ensure that our results are not biased or influenced by the choice of this measure, we also test the robustness of our results using alternative measures to approximate the productivity of firms (in the form of factor ratios). These are simple ratios of the corresponding variables. These measures of firm performance based on factor shares also have the advantage of being simple and very informative. In column (1), we repeat the estimation with our main variable (TFPR). In column (2), we estimate the effect of the power constraints on labor costs per USD of sales. The result suggests that the constraints induces an increase in production costs for the treated firms. Finally, in column (3), the constraints lead to a decrease in the amount of total sales per worker (in 2009 USD). In Table (17) in Appendix, we added additional potentially relevant control variables to minimize po- tential bias due to unobserved confounding factors. These variables are exports, size of the locality of firms, public ownership versus other actors’ ownership, gender of the top manager, quality certification, informal payment, investment in fixed assets, bank account and access to credit. These variables are potentially relevant variables in our model according to this literature. For each case, our results remain statistically significant. This further demonstrates the stability of our main results. 22 Table 10: Robustness tests – The effect of power outages on the firm’s revenue-based TFP and factor share based estimates of productivity (1) (2) (3) revenue-based TFP Cost of inputs per unit of sales Sales per labor cost Power Outages -0.102∗∗∗ 0.010∗∗ -1.193∗∗∗ (0.032) (0.005) (0.439) Observations 31403 31403 31403 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regres- sions (Entropy Balancing). The dependent variable is the logarithm of the firm-level revenue-based total factor productivity in regression (1), Cost of inputs per unit of sales in regression (2) and Sales per labor cost in regression (3). In each regression, the treatment variable represents power outages dummy variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentra- tion in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 5.4 Heterogeneity check We also tested the evolution of our results following the income group of the countries in which the firms operate and following the geographical regions. In Table (11), we have the effect of power constraints on TFPR following the income level of countries. We see that the negative effect found in Lower Income countries far exceeds those found in Middle Income countries. Indeed, the magnitude found in these Lower Income countries is 2.5 times that found in Upper Middle Income countries and 4.5 times that found in Lower Middle Income countries. For the Lower Income countries, this negative effect could be explained by the low level of resilience for firms in these fragile environments. For the Upper Middle Income countries, part of the explanation could lie in the fact that firms in these countries are relatively more energy-intensive, and therefore more exposed to energy constraints. The absence of a significant effect in the Middle Income countries would reflect the fact that firms in these countries are relatively more resilient than those in the Lower Income countries, and energy intensity is lower than in the Upper Middle Income countries. Table 11: Heterogeneity check – The effect of power outages on the firm’s revenue-based TFP by countries income group (1) (2) (3) Low Income Lower Middle Income Upper Middle Income Power Outages -0.214∗∗ -0.047 -0.085∗ (0.082) (0.031) (0.044) Observations 6848 14215 10340 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The dependent variable is the loga- rithm of the firm-level revenue-based total factor productivity for all regressions (1-3). In each regression, the treatment variable represents power outages dummy variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, For- eign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 In the Table (12), we then provide further evidence that the overall negative effect is particularly driven by Sub-Saharan African (column 1) and MENA countries (column 5). This highlights to some extent the 23 low resilience of firms in these regions (lack of alternative solutions such as back-up generators, low level of investment in R&D). Although the effect is negative in all regions, it remains statistically non-significant in the other regions. Firms in South Asia (SAR) remain the most resilient, followed by those in Europe and Central Asia (ECA). Table 12: Heterogeneity check – The effect of power outages on the firm’s revenue-based TFP by geograph- ical regions (1) (2) (3) (4) (5) (6) AFR EAP ECA LAC MNA SAR Power Outages -0.223∗∗∗ -0.052 -0.028 -0.082 -0.115∗∗∗ -0.026 (0.052) (0.092) (0.059) (0.050) (0.026) (0.066) Observations 7714 3586 3132 9120 3168 4683 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The dependent variable is the loga- rithm of the firm-level revenue-based total factor productivity for all regressions (1- 6). In each regression, the treatment variable represents power outages dummy vari- able. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (an- nual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate in- cludes Year FE, Country FE, Income group FE and a Constant. Sub-Saharan Africa (AFR), East Asia and Pacific (EAP), Europe and Central Asia (ECA), Latin Amer- ica and Caribbean (LAC), Middle East and North Africa (MNA), South Asia (SAR). Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 5.5 Transmission channels Our intuition is that power constraints directly affect the operations of firms by influencing their operating regime (capacity utilization) and inducing losses related to power outages (losses in % of sales). Column (1) in Table (13) indicates that power constraints significantly reduce the capacity utilization of the firms affected. This is linked to the fact that, in the event of a outage or a rise in the cost of kwh, firms can no longer operate at full capacity, so they first seek to readjust and reorganize their operating regime. Indeed, one of the consequences of the increase in power costs is a decrease in the size of the firm and productivity (Allcott et al., 2016), and therefore a lower profitability for the affected firm. Similarly, power outages, for example, cause sizeable damage to firms. This leads to considerable losses (increased losses) especially for firms that are not very resilient and do not have alternative measures such as back-up generators (column 2). 5.6 Mitigating factors Many developing countries are not able to provide their industrial sectors with reliable power, so many firms have to deal with an insufficient and unreliable power supply (Alby et al., 2013). The response of firms to unreliable power supply can vary. In its simplest form, it means additional costs if, for example, the firm has to buy and run a back-up generator, which also results in a higher unit cost of power (Elliott et al., 2021). It has to be said that affected firms often opt for self-generation of power, even though this is widely considered to be a second-best solution (Abdisa, 2018). Indeed, in Africa, for example, self-generated power is on average 313% more expensive than power from the grid (Alby et al., 2013). Are back-up 24 Table 13: Main channels – The effect of power outages on the firm’s Capacity Utilization (%) and Losses due to Outages (% of annual Sales) (1) (2) Capacity Utilization (%) Losses (% Sales) Power Outages -1.390∗∗ 6.680∗∗∗ (0.692) (0.700) Observations 30711 14397 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The dependent vari- able is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-2). In each regression, the treatment variable represents power outages dummy variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 generators really a solution to the constraints in the power sector in developing countries? In column (3) and (4) of the Table (14), we can see that having a back-up generator reduces the negative effect of power constraints by half, even if it does not fully correct the shock (Abdisa, 2020). Given the prominent role of research and development (R&D) in the operation of firms in general, we also tested a potential mitigating role for the R&D investments of the firms in our analysis. Through R&D, firms can mitigate the effect of unreliable energy supply by switching to less energy-intensive technologies (Alam, 2013), or can replace power with other types of fuels (Allcott et al., 2016) or materials (Fisher- Vanden et al., 2015). R&D can also instruct firms on the production of energy-intensive intermediates to be externalized instead of producing them internally. R&D also allows firms to effectively modify their production strategy. This ability to re-optimize decisions can therefore limit the negative effects of poor quality of power service for the affected firms (Alam, 2013). This is why we were also keen to test the mitigating role of R&D in column (1) and (2). The results suggest that investments in R&D allow firms to mitigate the negative effects of the constraints encountered in the power sector. Table 14: Mitigating factors – The effect of power outages on the firm’s revenue-based TFP (1) (2) (3) (4) No R&D R&D No Generator Generator Power Outages -0.151∗∗∗ -0.093 -0.179∗∗∗ -0.098∗∗ (0.034) (0.093) (0.035) (0.042) Observations 16152 5683 20229 10210 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The depen- dent variable is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-4). In each regression, the treatment variable represents power outages dummy variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Qual- ity). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 25 5.7 Placebo tests Finally, we perform some placebo tests to ensure that the effect captured is essentially that of the constraints observed in the power sector. To do this, in Table (15), we have dummy variables (1/2 1/2), (1/3 2/3), (2/3 1/3), (1/4 3/4) and (3/4 1/4). When replacing each of these random dummies in our baseline model, we find no statistically significant effect on firm-level revenue-based total factor productivity. Table 15: Placebo tests – The non-significant effects of random treatment variables on the firm’s revenue- based TFP (1) (2) (3) (4) (5) Random treatment (1/2 1/2) 0.012 (0.021) Random treatment (1/3 2/3) -0.002 (0.021) Random treatment (2/3 1/3) -0.025 (0.022) Random treatment (1/4 3/4) -0.006 (0.017) Random treatment (3/4 1/4) 0.030 (0.019) Observations 31406 31404 31404 31405 31404 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The dependent vari- ables is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-5). In each regression, the treatment variable represents a random treatment variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank con- centration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 6 Conclusion The objective of this paper was to analyze the role of power sector constraints on the performance level of firms in developing countries. Theoretically, we have shown that it is relatively complex to conclude the expected overall effect. Indeed, the constraints on the power sector lead to two types of effects in op- posite directions on the productivity of firms, namely a direct negative effect and an indirect mitigating effect (output prices adjustment producers). Empirically, our approach based on Entropy Balancing to con- trol endogeneity shows that the overall effect of power outages on revenue-based total factor productivity (TFPR) is negative and statistically significant. Indeed, we observed a 9% lower unconditional average pro- ductivity for exposed firms compared to non-exposed firms. Moreover, we establish a robust link between the severity of self-reported power constraints or obstacles by firms (minor, moderate, major, severe, and biggest) and the magnitude of productivity loss for firms. In other words, the greater the level of power constraints self-reported by a firm, the greater the effect of these constraints on its productivity, and vice versa. Our results are robust to changes in the level of clustering, the estimation model, and the measure of our productivity variable. We also show that the most affected firms operate mainly in Sub-Saharan Africa and in the MENA 26 region. Our results suggest that power constraints affect the TFPR by reducing firms’ capacity utilization (they no longer operate at full capacity) and by increasing direct losses due to numerous power outages. Finally, we have identified the acquisition of back-up generators and R&D investments as important factors in mitigating power constraints for firms in developing countries. In terms of recommendations, we would like first to recall that manufacturing firms are the main drivers of economic growth in developing countries. To fully play their part in the path to emergence, these firms need to benefit from a modern, reliable and affordable power service. However, this requires the estab- lishment of a financially viable power sector. To achieve this, on the supply side, the authorities must set up independent regulatory agencies in order to put in place appropriate measures, notably tariff (automatic tariff adjustment mechanisms, cost reflectivity, etc.) and non-tariff measures (development of a master plan for instance) to reassure investors. Combined with the opening-up of the power generation to private actors (through IPPs, PPPs, etc.), this will make it possible to move capital into the sector and thus reduce the investment gap. In parallel with this effort to improve the power infrastructure, strengthen the quality of the country’s institutions as a whole in order to attract foreign investors in various sectors and boost economic growth (Acemoglu et al., 2002; Rodrik, 2006) in order to drive the demand for power for the viability of the entire power sector. A viable power sector attracts more capital and provides quality power service to firms, which will in turn be more productive and competitive. This would create a virtuous circle for the whole economy. References Abdisa, L. T. (2018). Power outages, economic cost, and firm performance: Evidence from ethiopia. Utilities Policy, 53:111–120. Abdisa, L. T. (2020). Role of investment in self-generation in mitigating outage loss: Evidence from sub- saharan african firms. Energy, Ecology and Environment, 5(6):407–420. Abeberese, A. B., Ackah, C. G., and Asuming, P. O. (2021). Productivity losses and firm responses to electricity shortages: Evidence from ghana. The World Bank Economic Review, 35(1):1–18. Acemoglu, D., Johnson, S., and Robinson, J. A. (2002). Reversal of fortune: Geography and institutions in the making of the modern world income distribution. The Quarterly journal of economics, 117(4):1231– 1294. Agostino, M., Di Tommaso, M. R., Nifo, A., Rubini, L., and Trivieri, F. (2020). Institutional quality and firms’ productivity in european regions. Regional Studies, 54(9):1275–1288. Alam, M. (2013). Coping with blackouts: Power outages and firm choices. Department of Economics, Yale University. Alby, P., Dethier, J.-J., and Straub, S. (2013). Firms operating under electricity constraints in developing countries. The World Bank Economic Review, 27(1):109–132. Allcott, H., Collard-Wexler, A., and O’Connell, S. D. (2016). How do electricity shortages affect industry? evidence from india. American Economic Review, 106(3):587–624. 27 Apeti, A. E. (2023a). Household welfare in the digital age: Assessing the effect of mobile money on household consumption volatility in developing countries. World Development, 161:106110. Apeti, A. E. (2023b). Sovereign debt default and inequality. Industrial and Corporate Change. Apeti, A. E., Bambe, B.-W.-W., and Combes, J.-L. (2023). On the Macroeconomic Effects of Fiscal Re- forms: Fiscal Rules and Public Expenditure Efficiency. hal-03961062f. Apeti, A. E. and Edoh, E. D. (2023). Tax revenue and mobile money in developing countries. Journal of Development Economics, 161:103014. Asiedu, E., Azomahou, T. T., Gaekwad, N. B., and Ouedraogo, M. (2021). The determinants of electricity constraints by firms in developing countries. Energy Economics, 104:105605. Asiedu, E. and Freeman, J. (2009). The effect of corruption on investment growth: Evidence from firms in latin america, sub-saharan africa, and transition countries. Review of Development Economics, 13(2):200–214. Bambe, B.-W.-W., Combes, J. L., Kaba, K., and Minea, A. (2022). Inflation targeting and developing countries’ performance: Evidence from firm-level data. Technical report, Orleans Economics Labora- tory/Laboratoire d’Economie d’Orleans (LEO . . . . Barth III, M. J. and Ramey, V. A. (2001). The cost channel of monetary transmission. NBER macroeco- nomics annual, 16:199–240. Bernanke, B. S. (1983). Irreversibility, uncertainty, and cyclical investment. The quarterly journal of economics, 98(1):85–106. Bodenstein, M., Erceg, C. J., and Guerrieri, L. (2011). Oil shocks and external adjustment. Journal of International Economics, 83(2):168–184. Brandt, L., Tombe, T., and Zhu, X. (2013). Factor market distortions across time, space and sectors in china. Review of Economic Dynamics, 16(1):39–58. Canan, C., Lesko, C., and Lau, B. (2017). Instrumental variable analyses and selection bias. Epidemiology (Cambridge, Mass.), 28(3):396. Chauvet, L. and Ehrhart, H. (2018). Aid and growth: evidence from firm-level data. Journal of Development Economics, 135:461–477. Chauvet, L. and Ferry, M. (2021). Taxation, infrastructure, and firm performance in developing countries. Public Choice, 187:455–480. Chauvet, L. and Jacolin, L. (2017). Financial inclusion, bank concentration, and firm performance. World Development, 97:1–13. Cheng, D., Shi, X., Yu, J., and Zhang, D. (2019). How does the chinese economy react to uncertainty in international crude oil prices? International Review of Economics & Finance, 64:147–164. 28 Choi, S., Furceri, D., Loungani, M. P., Mishra, M. S., and Poplawski-Ribeiro, M. (2017). Oil prices and inflation dynamics: evidence from advanced and developing economies. International Monetary Fund. Cole, M. A., Elliott, R. J., Occhiali, G., and Strobl, E. (2018). Power outages and firm performance in sub-saharan africa. Journal of Development Economics, 134:150–159. Elder, J. and Serletis, A. (2009). Oil price uncertainty in canada. Energy Economics, 31(6):852–856. Elliott, R. J., Nguyen-Tien, V., and Strobl, E. A. (2021). Power outages and firm performance: A hydro-iv approach for a single electricity grid. Energy economics, 103:105571. Ertefaie, A., Small, D., Flory, J., and Hennessy, S. (2016). Selection bias when using instrumental variable methods to compare two treatments but more than two treatments are available. The international journal of biostatistics, 12(1):219–232. 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. Foster, L., Haltiwanger, J., and Syverson, C. (2008). Reallocation, firm turnover, and efficiency: Selection on productivity or profitability? American Economic Review, 98(1):394–425. Grainger, C. A. and Zhang, F. (2017). The impact of electricity shortages on firm productivity: Evidence from pakistan. World Bank Policy Research Working Paper, (8130). Hainmueller, J. (2012). Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. Political analysis, 20(1):25–46. Halvorsen, R., Palmquist, R., et al. (1980). The interpretation of dummy variables in semilogarithmic equations. American economic review, 70(3):474–475. Hardy, M. and McCasland, J. (2021). Lights off, lights on: The effects of electricity shortages on small firms. The World Bank Economic Review, 35(1):19–33. Hsieh, C.-T. and Klenow, P. J. (2009). Misallocation and manufacturing tfp in china and india. The Quar- terly journal of economics, 124(4):1403–1448. Jones, C. I. (2011). Misallocation, economic growth, and input-output economics. Technical report, Na- tional bureau of economic research. Lee, C.-C., Wang, C.-W., and Ho, S.-J. (2020). Financial inclusion, financial innovation, and firms’ sales growth. International Review of Economics & Finance, 66:189–205. Leung, D., Meh, C., and Terajima, Y. (2008). Firm size and productivity. Technical report, Bank of Canada. Levinsohn, J. and Petrin, A. (2003). Estimating production functions using inputs to control for unobserv- ables. The review of economic studies, 70(2):317–341. Lin, S. and Ye, H. (2009). Does inflation targeting make a difference in developing countries? Journal of Development economics, 89(1):118–123. 29 Majumdar, S. K. (1997). The impact of size and age on firm-level performance: some evidence from india. Review of industrial organization, 12(2):231–241. Moulton, B. R. (1990). An illustration of a pitfall in estimating the effects of aggregate variables on micro units. The review of Economics and Statistics, pages 334–338. Neuenkirch, M. and Neumeier, F. (2016). The impact of us sanctions on poverty. Journal of Development Economics, 121:110–119. Prescott, E. C. (2002). Prosperity and depression. American Economic Review, 92(2):1–15. Rodrik, D. (2006). Goodbye washington consensus, hello washington confusion? a review of the world bank’s economic growth in the 1990s: learning from a decade of reform. Journal of Economic literature, 44(4):973–987. Sadorsky, P. (1999). Oil price shocks and stock market activity. Energy economics, 21(5):449–469. Scott, A., Darko, E., Lemma, A., and Rud, J.-P. (2014). How does electricity insecurity affect businesses in low and middle income countries. Shaping policy for development, pages 1–80. Shi, X. and Sun, S. (2017). Energy price, regulatory price distortion and economic growth: a case study of china. Energy Economics, 63:261–271. age et de la taille sur la performance financi` St-Pierre, J., Julien, P.-A., and Morin, M. (2010). L’effet de l’ˆ ere ´ conomique des pme. Journal of Small Business & Entrepreneurship, 23(2):287–306. et e Syverson, C. (2011). What determines productivity? Journal of Economic literature, 49(2):326–65. Tovar, B., Ramos-Real, F. J., and De Almeida, E. F. (2011). Firm size and productivity. evidence from the electricity distribution industry in brazil. Energy Policy, 39(2):826–833. Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data. MIT press. Xiao, Z., Gao, J., Wang, Z., Yin, Z., and Xiang, L. (2022). Power shortage and firm productivity: Evidence from the world bank enterprise survey. Energy, 247:123479. 30 A Appendix Table 16: Main variables description Variables Description Source TFPR TFPR (VAKL model) World Bank Group, Enter- prise Analysis Unit. 2017. “firm-level Productivity Es- timates”. Cost Cost of inputs per unit of sales – Sales Sales per labor cost – Capacity Capacity Utilization (%) The World Bank Enterprise Surveys (WBES) R&D Investmet in R&D – Generator Generator acquisition by the firm – Losses Losses due to Outages (% of annual Sales) – Outages Firm facing Power Outages – Obstacles How Much Of An Obstacle: Electricity To Operations – Of This Establishment? Biggest Power as biggest obstacle – Size Firm size – Age Firm’s longevity – Private Foreign private participation – Website Own website – Regulation Regulatory Quality Worldwide Governance In- dicators - World Bank Data- Bank Bank Bank concentration in %: Assets of three largest com- Bankscope and Orbis Bank mercial banks as a share of total commercial banking Focus, Bureau van Dijk assets. Total assets include total earning assets, cash (BvD) and due from banks, foreclosed real estate, fixed assets, goodwill, other intangibles, current tax assets, deferred tax assets, discontinued operations and other assets. ODA Net official development assistance (ODA) received in World Development Indica- % of GNI: tors — DataBank Inflation Inflation, consumer prices (annual %) – Growth GDP growth (annual %) – Capita GDP per capita in USD – FINDEX Financial development index Global Financial Develop- ment Database - World Bank Exports Direct Exports (% of Sales) The World Bank Enterprise Surveys (WBES) Locality Size Of Locality – Government % Owned By Government/State – Other % Owned By Other – Female Is The Top Manager Female? – Certification Does Establishment Have An Internationally- – Recognized Quality Certification? Informal Percent Of Total Annual Sales Paid In Informal Pay- – ments Investment Did This Establishment Purchase Any Fixed Assets In – Last Fiscal Yr? Account Does This Establishment Have A Checking AndSaving – Account? Credit Establishment Has A Line Of Credit Or Loan From A – Financial Institution? i Table 17: Robustness tests – The effect of power outages on the firm’s revenue-based TFP by including additional control variables to the baseline specification (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Power Outages -0.102∗∗∗ -0.090∗∗∗ -0.086∗∗ -0.091∗∗∗ -0.091∗∗∗ -0.078∗∗ -0.091∗∗∗ -0.087∗∗∗ -0.095∗∗∗ -0.087∗∗∗ -0.090∗∗∗ (0.032) (0.029) (0.035) (0.029) (0.029) (0.035) (0.029) (0.031) (0.029) (0.028) (0.029) Direct Exports (% of Sales) 0.001∗ (0.001) Size Of Locality -0.068∗∗∗ (0.023) % Owned By Government/State -0.000 (0.003) % Owned By Other 0.002∗∗ (0.001) Female Top Manager -0.017 (0.024) Recognized Quality Certification 0.007 (0.008) Informal Payments (% Of Sales) 0.003 (0.002) Purchase of Fixed Assets -0.048∗∗∗ (0.012) Checking/Saving Account -0.011 (0.013) Line Of Credit Or Loan 0.014 (0.008) Observations 31403 31402 25217 31403 31403 22999 31397 25342 31402 30448 31401 Notes: Table shows the average treatment effects on the treated obtained by weighted least squares regressions (Entropy Balancing). The dependent variable is the logarithm of the firm-level revenue-based total factor productivity for all regressions (1-11). In each regression, the treatment variable represents power outages dummy variable. Firm-level controls variables are: Firm size, Firm’s longevity, Own website, Foreign private participation. Country-level controls variables are: GDP growth (annual %), Inflation (annual %), log(Bank concentration in %), log(FD index), log(GDP per capita), log(Net ODA in % of GNI), log(Regulatory Quality). Each row represents an additional control variable. Each estimate includes Year FE, Country FE, Region FE, Income group FE and a Constant. Standard errors in parentheses are clustered at country level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Table 18: Summary statistics on revenue-based TFP by outages and over regions Mean Std. dev. Min Max N No Outages AFR 2.76 1.49 -1.47 8.58 1504.00 EAP 2.47 1.46 -1.89 8.11 1535.00 ECA 2.12 1.58 -2.13 8.51 2015.00 LAC 2.41 1.41 -2.44 7.41 4271.00 MNA 2.59 1.51 -2.13 8.45 1553.00 SAR 2.54 1.58 -1.79 7.53 1439.00 Outages AFR 2.63 1.44 -1.85 8.04 6210.00 EAP 2.32 1.46 -1.76 8.83 2051.00 ECA 2.22 1.64 -2.52 8.16 1117.00 LAC 2.34 1.46 -3.23 7.54 4849.00 MNA 2.53 1.53 -1.92 7.24 1615.00 SAR 2.50 1.56 -1.86 8.29 3244.00 Table 19: Summary statistics on capacity utilization by outages and over regions Mean Std. dev. Min Max Obs. No outages AFR 73.74 22.80 0.00 100.00 1469.00 EAP 82.31 21.05 0.00 100.00 1498.00 ECA 71.57 24.31 0.00 100.00 1958.00 LAC 72.20 20.70 0.00 100.00 4192.00 MNA 67.99 23.11 1.00 100.00 1487.00 SAR 82.59 16.79 10.00 100.00 1435.00 Outages AFR 71.08 20.56 1.00 100.00 6075.00 EAP 79.34 21.17 1.00 100.00 2017.00 ECA 68.95 24.12 1.00 100.00 1088.00 LAC 71.39 20.47 0.00 100.00 4709.00 MNA 71.91 21.38 1.00 100.00 1556.00 SAR 76.79 16.21 2.00 100.00 3227.00 ii Table 20: Summary statistics on the number of observations by region and over years of survey Region AFR EAP ECA LAC MNA SAR Total Year of survey 2006 1,416 3,254 4,670 2007 3,434 204 3,638 2008 303 303 2009 257 1,523 601 686 109 3,176 2010 117 3,945 4,062 2011 27 230 257 2012 28 28 2013 936 57 536 1,910 1,423 4,862 2014 364 179 2,863 3,406 2015 232 630 58 920 2016 227 999 282 761 2,269 2017 72 953 1,025 2018 519 66 237 822 2019 113 104 1,251 497 1,965 Total 7,714 3,586 3,132 9,120 3,168 4,683 31,403 Table 21: Summary statistics on the number of observations by region and by income group Region AFR EAP ECA LAC MNA SAR Total WB income group Low Income 5,381 111 1,356 6,848 Lower Middle Income 1,433 3,098 1,184 2,774 2,399 3,327 14,215 Upper Middle Income 900 488 1,837 6,346 769 10,340 Total 7,714 3,586 3,132 9,120 3,168 4,683 31,403 Table 22: Summary statistics on the number of observations by industry and by region Region AFR EAP ECA LAC MNA SAR Total Industry stratification Basic Metals & Metal Products 83 300 383 Basic Metals/Fabricated Metals/Machinery & Equip. 83 398 481 Chemicals & Chemical Products 8 132 588 184 326 1,238 Chemicals, Non-Metallic Mineral, Plastics & Rubber 38 38 Chemicals, Plastics & Rubber 97 367 464 Electronics 63 99 162 Electronics & Communications Equip. 62 209 271 Fabricated Metal Products 68 154 181 183 65 240 891 Food 1,768 413 442 1,726 662 515 5,526 Furniture 116 227 92 81 516 Garments 906 438 227 696 340 266 2,873 Leather Products 75 119 100 294 Machinery & Equipment 165 302 279 746 Machinery & Equipment, Electronics & Vehicles 49 49 Manufacturing 1,524 1,067 1,539 1,111 68 328 5,637 Manufacturing Panel 62 62 Metals, Machinery, Computer & Electronics 57 57 Minerals, Metals, Machinery & Equipment 55 55 Mining Related Manufacturing 15 15 Motor Vehicles 61 188 249 Motor Vehicles & Transport Equip. 29 29 Non-Metallic Mineral Products 68 302 163 229 260 220 1,242 Other Manufacturing 2,367 541 415 995 673 1,094 6,085 Petroleum products, Plastics & Rubber 89 89 Printing & Publishing 50 33 83 Rest of Universe 358 901 1,259 Rubber & Plastics Products 261 128 86 255 730 Textiles 24 153 222 152 253 804 Textiles & Garments 143 719 83 945 Wood Products 57 57 Wood products, Furniture, Paper & Publishing 73 73 Total 7,714 3,586 3,132 9,120 3,168 4,683 31,403 iii Table 23: Summary statistics on the number of observations by industry and over years of survey Year of survey 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 Total Industry stratification Basic Metals & Metal Products 300 83 383 Basic Metals/Fabricated Metals/Machinery & Equip. 398 83 481 Chemicals & Chemical Products 396 8 173 118 219 213 33 78 1,238 Chemicals, Non-Metallic Mineral, Plastics & Rubber 38 38 Chemicals, Plastics & Rubber 367 47 50 464 Electronics 99 36 27 162 Electronics & Communications Equip. 209 62 271 Fabricated Metal Products 42 103 65 141 65 264 89 122 891 Food 1,173 910 33 338 607 83 765 320 161 309 200 234 393 5,526 Furniture 70 84 110 194 25 33 516 Garments 656 724 39 322 106 72 474 17 120 120 52 171 2,873 Leather Products 75 162 57 294 Machinery & Equipment 131 47 83 88 22 279 96 746 Machinery & Equipment, Electronics & Vehicles 49 49 Manufacturing 690 163 914 886 27 28 838 241 810 342 119 579 5,637 Manufacturing Panel 62 62 Metals, Machinery, Computer & Electronics 57 57 Minerals, Metals, Machinery & Equipment 55 55 Mining Related Manufacturing 15 15 Motor Vehicles 61 188 249 Motor Vehicles & Transport Equip. 29 29 Non-Metallic Mineral Products 229 223 193 246 113 105 133 1,242 Other Manufacturing 856 736 21 464 615 75 1,309 546 309 138 273 272 471 6,085 Petroleum products, Plastics & Rubber 89 89 Printing & Publishing 39 33 11 83 Rest of Universe 901 358 1,259 Rubber & Plastics Products 112 128 86 255 35 114 730 Textiles 149 226 199 230 804 Textiles & Garments 443 72 33 149 210 38 945 Wood Products 57 57 Wood products, Furniture, Paper & Publishing 73 73 Total 4,670 3,638 303 3,176 4,062 257 28 4,862 3,406 920 2,269 1,025 822 1,965 31,403 Table 24: Sub-Saharan Africa (AFR) WB income group Low Income Lower Middle Income Upper Middle Income Total Official Country Name Angola 206 206 Benin 53 53 Botswana 158 158 Burkina Faso 26 26 Burundi 134 134 Cameroon 120 120 Congo, Dem. Rep. 298 298 Eswatini 64 64 Ethiopia 259 259 Gambia, The 83 83 Ghana 278 186 464 Guinea 5 5 Kenya 600 261 861 Lesotho 39 39 Liberia 66 66 Madagascar 230 230 Mali 350 350 Mauritania 77 13 90 Mauritius 57 57 Mozambique 532 532 Namibia 94 13 107 Niger 6 6 Nigeria 930 216 1,146 Rwanda 171 171 Senegal 250 86 336 South Africa 672 672 Tanzania 362 362 Togo 19 19 Uganda 353 353 Zambia 299 148 447 Total 5,381 1,433 900 7,714 iv Table 25: Sub-Saharan Africa (AFR) Year of survey 2006 2007 2009 2010 2011 2013 2014 2015 2016 2017 2018 2019 Total Official Country Name Angola 199 7 206 Benin 53 53 Botswana 110 48 158 Burkina Faso 26 26 Burundi 98 36 134 Cameroon 55 65 120 Congo, Dem. Rep. 143 49 106 298 Eswatini 64 64 Ethiopia 27 232 259 Gambia, The 30 53 83 Ghana 278 186 464 Guinea 5 5 Kenya 387 213 261 861 Lesotho 39 39 Liberia 66 66 Madagascar 119 111 230 Mali 291 13 46 350 Mauritania 77 13 90 Mauritius 57 57 Mozambique 327 205 532 Namibia 94 13 107 Niger 6 6 Nigeria 930 216 1,146 Rwanda 58 113 171 Senegal 250 86 336 South Africa 672 672 Tanzania 254 108 362 Togo 19 19 Uganda 289 64 353 Zambia 299 148 447 Total 1,416 3,434 257 117 27 936 364 232 227 72 519 113 7,714 Table 26: East Asia and Pacific (EAP) WB income group Lower Middle Income Upper Middle Income Total Official Country Name Cambodia 109 109 Indonesia 529 529 Lao PDR 198 198 Mongolia 283 283 Myanmar 477 477 Philippines 579 579 Thailand 488 488 Vietnam 923 923 Total 3,098 488 3,586 Table 27: East Asia and Pacific (EAP) Year of survey 2009 2012 2013 2014 2015 2016 2018 2019 Total Official Country Name Cambodia 109 109 Indonesia 529 529 Lao PDR 28 104 66 198 Mongolia 122 57 104 283 Myanmar 179 298 477 Philippines 324 255 579 Thailand 488 488 Vietnam 548 375 923 Total 1,523 28 57 179 630 999 66 104 3,586 v Table 28: Europe and Central Asia (ECA) WB income group Low Income Lower Middle Income Upper Middle Income Total Official Country Name Albania 25 78 103 Armenia 81 81 Azerbaijan 92 92 Belarus 324 324 Bosnia and Herzegovina 168 168 Croatia 179 179 Georgia 101 94 195 Kazakhstan 555 555 Kyrgyz Republic 54 107 161 Moldova 159 159 North Macedonia 133 133 Serbia 214 214 Tajikistan 57 57 Ukraine 711 711 Total 111 1,184 1,837 3,132 Table 29: Europe and Central Asia (ECA) Year of survey 2007 2008 2009 2013 2018 2019 Total Official Country Name Albania 25 14 64 103 Armenia 58 23 81 Azerbaijan 70 8 14 92 Belarus 40 47 237 324 Bosnia and Herzegovina 64 60 44 168 Croatia 179 179 Georgia 66 35 94 195 Kazakhstan 102 35 418 555 Kyrgyz Republic 54 25 82 161 Moldova 90 13 56 159 North Macedonia 60 73 133 Serbia 103 55 56 214 Tajikistan 57 57 Ukraine 140 148 423 711 Total 204 303 601 536 237 1,251 3,132 Table 30: Latin America and Caribbean (LAC) WB income group Lower Middle Income Upper Middle Income Total Official Country Name Bolivia 303 303 Brazil 686 686 Chile 1,024 1,024 Colombia 946 946 Costa Rica 195 195 Dominican Republic 96 96 Ecuador 250 187 437 El Salvador 597 597 Guatemala 484 79 563 Guyana 57 57 Honduras 321 321 Jamaica 80 80 Mexico 1,760 1,760 Nicaragua 321 321 Panama 21 21 Paraguay 201 39 240 Peru 240 792 1,032 St. Lucia 47 47 Suriname 73 73 Uruguay 321 321 Total 2,774 6,346 9,120 vi Table 31: Latin America and Caribbean (LAC) Year of survey 2006 2009 2010 2016 2017 Total Official Country Name Bolivia 215 35 53 303 Brazil 686 686 Chile 435 589 1,024 Colombia 540 406 946 Costa Rica 195 195 Dominican Republic 76 20 96 Ecuador 250 88 99 437 El Salvador 310 77 210 597 Guatemala 273 211 79 563 Guyana 57 57 Honduras 196 73 52 321 Jamaica 80 80 Mexico 798 962 1,760 Nicaragua 249 72 321 Panama 21 21 Paraguay 131 70 39 240 Peru 240 515 277 1,032 St. Lucia 47 47 Suriname 73 73 Uruguay 157 164 321 Total 3,254 686 3,945 282 953 9,120 Table 32: Middle East and North Africa (MNA) WB income group Lower Middle Income Upper Middle Income Total Official Country Name Djibouti 2 2 Egypt, Arab Rep. 2,057 2,057 Jordan 210 210 Lebanon 330 330 Morocco 340 340 Tunisia 229 229 Total 2,399 769 3,168 Table 33: Middle East and North Africa (MNA) Year of survey 2013 2016 2019 Total Official Country Name Djibouti 2 2 Egypt, Arab Rep. 1,296 761 2,057 Jordan 183 27 210 Lebanon 106 224 330 Morocco 94 246 340 Tunisia 229 229 Total 1,910 761 497 3,168 Table 34: South Asia (SAR) WB income group Low Income Lower Middle Income Total Official Country Name Bangladesh 1,028 1,028 Bhutan 58 58 India 2,863 2,863 Nepal 328 328 Pakistan 176 176 Sri Lanka 230 230 Total 1,356 3,327 4,683 vii Table 35: South Asia (SAR) Year of survey 2009 2011 2013 2014 2015 Total Official Country Name Bangladesh 1,028 1,028 Bhutan 58 58 India 2,863 2,863 Nepal 109 219 328 Pakistan 176 176 Sri Lanka 230 230 Total 109 230 1,423 2,863 58 4,683 viii