## wp18159

## Source details

**Canonical URL:** [wp18159](https://www.imf.org/-/media/files/publications/wp/2018/wp18159.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2018/wp18159.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2018/wp18159.pdf.json)

---

### 3.1 Conceptual Framework — power sector, firm choices, and measurement
- Power sector structure and firm choice environment
  - Power generation and distribution in Ethiopia is entrusted to the state utility company; customers do not have the option to choose alternative power suppliers.
  - Absent retail choice, firms facing power disruptions adopt coping mechanisms or absorb the full cost of outages.
  - Decision tree outcomes (summary):
    - If electricity disruptions are not a major constraint → Business as usual.
    - If disruptions are a major constraint → Coping mechanisms or shutdown.
      - Coping mechanisms:
        - Self-generate → ties up potentially productive capital K; may provide temporary relief but can reduce long-run output; self-generation does not guarantee complete substitution between grid and self-generated power.
        - Reduce capacity utilization (idle capacity) → small productive K; high cost of self-generation.
      - Shutdown → temporary cessation of production (idle capacity).
- Economic implications of self-generation and heterogeneous firm responses
  - Self-generation:
    - Allows replacement of grid electricity and continued production throughout the year for firms that can acquire generators.
    - Installation of self-generation capacity ties up potentially productive capital K.
    - The capacity to self-generate does not automatically guarantee complete substitution between grid and self-generated power.
    - The economic burden of self-generation may not be wholly offloaded to consumers through mark-ups if:
      - Firms face elastic demand for their goods.
      - Industry-specific tariff rates are low enough to allow cheaper imports.
    - Self-generation increases capital and operating costs, raises production costs, lowers competitiveness of local products, and limits potential for exploiting economies of scale (Steinbuks and Foster, 2010).
  - Heterogeneity by firm size:
    - Self-generation option is often open only to medium and large firms.
    - Majority of smaller firms either:
      - Have small-scale operations making it too costly to own generators, or
      - Are rationed out of credit markets that facilitate such investments.
    - Small firms facing unanticipated disruptions are forced to scale down: lower intermediate input demand, vary capacity utilization over time, or undertake temporary shutdown.
- Measurement of power disruptions used in analysis
  - Two alternative measures:
    - S: dummy equal to one if the firm faced electricity disruption in a fiscal year (self-reported). Distinguishes ‘shortage-firms’ from ‘non-shortage firms’. The designation is time-varying.
      - Cautions:
        - Self-reported measures can underestimate the proportion of firms experiencing disruptions if generator-owning firms are less likely to report disruptions.
        - Timing of judgments about major constraints can induce endogeneity if firms decide on constraints after observing their productivities.
    - Sσ: dummy capturing all firms facing grid-power consumption drops exceeding ‘1standard deviation’ from long-run grid power consumption in any fiscal year (includes firms that cope or incur full cost).
      - Data shows Sσ is not correlated with other distortions that might lower grid-power consumption such as raw material and machinery-part shortages (see Appendix Table A.3).
  - Observed relationship between measures:
    - The share of firms reporting power disruptions as a major constraint grew from below15percent of LMMI during2000−2006 to nearly25percent in2008 and33 percent in 2009.
    - The share of firms with electricity consumption (logs) below their long-run levels also increased.
    - The two measures track one another towards the end of the sample but are generally weakly correlated.
    - Except for 2003, Sσ > S; the difference Sσ − S may approximate firms that cope with power disruptions.
- Conceptual link to productivity and shutdown analysis
  - Two long-term implications emphasized:
    - Creation and perpetuation of idle capacity.
    - Tying up resources in unproductive capital (self-generation).
  - These channels feed empirical analysis by:
    - Quantifying productivity cost of power disruptions across firms varying by power use and productivity levels.
    - Examining the link between recurring blackouts and firms’ decisions to temporarily shut down to avoid losses.
  - Productivity losses computed in the study are a sum total of all conceivable costs.

### 5.1 How costly are power disruptions? — productivity, non-linearity, and shutdowns
- Labor productivity losses — benchmark and robustness
  - Pooled OLS benchmark (Table 3) finds factor elasticities (log-labor units): capital 0.10***, raw materials 0.63***, electricity 0.07*** (robust s.e. in parentheses).
  - Power disruption dummy: shortage-firms have "14-18percent" lower labor productivity than non-shortage firms (Columns 2 and 3).
  - Pooled OLS average estimates of productivity loss: "17-18percent" (Table 3 OLS averages).
  - Fixed effects (FE) and IV comparisons (Table 4 and Appendix):
    - FE β̃k (capital) is much lower than OLS.
    - IV (lagged investment per labor as instrument for capital) yields higher implied losses from power disruption; Hausman test favors IV over OLS (FE), but IV suffers loss of information due to many zero-investment observations.
    - Power disruption dummies are negative but smaller under FE/IV: pooled OLS ~17-18percent vs FE/IV average productivity loss about "5-6percent".
- Non-linear relationship with electricity intensity
  - Extended specifications include electricity intensity ̃e and ̃e^2 and interactions with disruption dummy.
  - Squared term marginally significant in some specifications indicating non-linearity.
  - FE predicted productivity losses (Table 4 Column 4 marginal effects):
    - Average productivity losses of "6-7percent" at mean intensity.
    - "22-30percent" at low intensity.
    - "16-22percent" at high intensity.
  - Interpretation of inverted U-shaped loss curve (when S is disruption measure):
    - Firms at low intensity may find self-generation unaffordable and resort to strategic shutdowns.
    - Large/power-intensive firms may self-generate but incur high self-generation costs, producing higher losses.
  - Alternative disruption measure S_σ produces a different (downward sloping) loss curve, with firms at higher intensities facing sharper losses under that measure.
- Total factor productivity (TFP) losses — control-function approaches and quantile regressions
  - TFP estimation approaches used: Levinshohn and Petrin (LP) and Ackerberg et al. (ACF) control-function approaches.
  - Wide variance in estimated elasticities between LP and ACF when raw-material demand is the control function: LP labor and capital elasticities are about half the values of ACF in the raw-material control case.
  - Panel quantile regressions at the median (Table 5):
    - MCMC with 50,000 draws per quantile used to estimate Eq.7.
    - LP-based TFP median results indicate average productivity loss of "5-6percent" due to power disruptions.
    - ACF-based TFP median results indicate average productivity loss of "nearly9percent" of TFP.
    - Table 5 disruption coefficients (selected): LP columns 0.06*, -0.05**, -0.05*; ACF columns 0.02, -0.09***, -0.09*** (median, bootstrapped s.e. in parentheses).
  - Within-industry median results (Table 6):
    - LP within-industry point estimates indicate average productivity loss of "7percent" due to power disruptions.
    - ACF within-industry estimates show similar patterns though point estimates differ; disruption coefficients include -0.07*** and -0.03* in various columns.
  - Productivity loss across quantiles (Figure 5):
    - Inverted U-shaped productivity loss curve across productivity distribution for aggregate TFP.
    - Firms at lower productivity quantiles Q_10 experience about "6 percent" loss; upper quantiles incur about "7-8percent" TFP losses.
    - Shape of loss curve varies with disruption measure S_σ.
- Shutdown behavior as a coping response
  - Empirical strategy: panel count-outcome models (Negative Binomial and Poisson) for months of shutdown I_it over firm age (Age as exposure variable).
  - Descriptive patterns (Figure 6):
    - Plant shutdown episodes constitute "37percent" of the full sample (top-left panel).
    - Younger firms shutdown more often than mature firms.
  - Count-model results (Table 7 and Appendix Table A.12):
    - Random effects NB and conditional fixed effects NB show a positive and significant relationship between power disruptions and shutdown incidents.
    - On average, firms are likely to shutdown operations for "1.4-2months" in response to power disruptions.
    - Shutdown incidents increased by about "1.3months" when unexploited capacity in the past year was "33% higher".
    - Poisson alternative yields similar conclusions.
    - Productivity levels are not significant predictors of shutdown episodes across models and productivity concepts.
  - Additional robustness: differenced panel data models and alternative activity-period measure A = 12 − I are examined.

### Appendix and robustness — coefficients, panels, and loss magnitudes
- Appendix Table A.13 (differenced OLS) — effect on activity months
  - Power disruptions are likely to compel firms to shutdown for nearly 1months.
  - Results remain consistent when lagged second differences of productivity and excess capacity are used.
  - A 33% increase in underutilized capacity is equivalent to1natural-log point increase inD f. From see Eq11, the response in months is exp(γ s ).
- Selected Negative Binomial coefficients (key entries preserved)
  - S: 0.61*** (0.05); 0.62*** (0.1); 0.67*** (0.11); 0.67*** (0.10); 0.69*** (0.10); 0.68*** (0.08); 0.57*** (0.05); 0.56*** (0.10); 0.63*** (0.13); 0.59*** (0.11); 0.60*** (0.10); 0.60*** (0.11)
  - D F t−1: 0.27*** (0.05); 0.29*** (0.09); 0.23*** (0.06); 0.28*** (0.06); 0.23*** (0.06); 0.17*** (0.07); 0.13 (0.08); 0.14** (0.07); 0.15** (0.06); 0.15*** (0.06)
  - ln TFP LP1 t−1: -0.1* (0.09); -0.11 (0.08)
  - ln TFP ACF1 t−1: -0.05 (0.07); -0.01 (0.09)
  - Const.: -1.99*** (0.12); -2.61*** (0.17); -2.92*** (0.31); -2.74*** (0.19); -3.15*** (0.22); -2.90*** (0.19); -2.66*** (0.091); -3.10*** (0.16); -3.09*** (0.34); -3.26*** (0.22); -3.01*** (0.25); -3.70*** (0.20)
  - Note: bootstrapped s.e. in parenthesis.
- Sample sizes, firms, and log-likelihood (selected)
  - Obs.: 11,094; 6,431; 4,856; 6,040; 6,040; 6,044; 6,418; 3,353; 2,151; 3,071; 3,071; 3,075
  - Firms: 3,280; 1,927; 1,508; 1,789; 1,789; 1,790; 1,319; 778; 494; 689; 689; 690
  - Log-Lik: -18320; -9112; -5829; -8140; -8136; -8144; -8149; -3945; -2506; -3539; -3539; -3540
- Selected aggregate and industry-level production function estimates (LP and ACF)
  - Aggregate LP (Table A.7): l 0.29*** (0.01), k 0.16*** (0.02), m 0.60*** (0.01); Obs 8,372; Firms 2,684.
  - Aggregate ACF (Table A.7): l 0.67*** (0.02), k 0.30*** (0.01), m 0.64*** (0.01).
  - Industry examples (LP, Table A.8):
    - CGP: l 0.21*** (0.03), k 0.10*** (0.04), e 0.18*** (0.04); Obs 1,757; Firms 619
    - FAB: l 0.34*** (0.03), k 0.14*** (0.03), e 0.16*** (0.03); Obs 2,386; Firms 728
  - Industry examples (ACF, Table A.9):
    - CGP: l 0.63*** (0.06), k 0.43*** (0.07), e 0.28*** (0.04)
    - FAB: l 0.69*** (0.08), k 0.30*** (0.03), e 0.26*** (0.04)
- Estimated productivity losses from electricity disruptions — selected IV marginal effects by intensity (̃e)
  - Unbalanced panel (Table A.4, Col (8)): Productivity Loss (1 vs 0) 2: -0.52; 3: -0.42; 4: -0.32; 5: -0.23; 6: -0.16; 7: -0.13; 8: -0.12; 9: -0.15; 10: -0.21; 11: -0.29. Joint χ2 p val = 0.00 (Joint = 47.18).
  - Unbalanced panel (Table A.6, Col (6)): (1 vs 0) 2: 0.18; 3: -0.02; 4: -0.16; 5: -0.25; 6: -0.29; 7: -0.31; 8: -0.30; 9: -0.26; 10: -0.19; 11: -0.07. Joint χ2 p val = 0.00 (Joint = 68.60).
  - Balanced panel (Table A.16, Col (8)): (1 vs 0) 2: -0.53; 3: -0.37; 4: -0.22; 5: -0.10; 6: -0.03; 7: -0.04; 8: -0.10; 9: -0.23; 10: -0.39; 11: -0.54. Joint χ2 p val = 0.00 (Joint = 22.68).
  - Balanced panel (Table A.18, Col (8)): (1 vs 0) 2: 0.33; 3: 0.14; 4: 0.01; 5: -0.09; 6: -0.15; 7: -0.19; 8: -0.21; 9: -0.21; 10: -0.18; 11: -0.14. Joint χ2 p val = 0.00 (Joint = 15.79).
- Count and alternative models (Table A.12 and pooled OLS Table A.13)
  - Poisson/random effects findings: disruption coefficients consistently positive and significant (e.g., Random Effects Disruption 0.38*** (0.04), Obs 11,094; Firms 3,280; Log Lik -19331).
  - Pooled OLS (Table A.13): Disruption coefficients -0.46*** to -0.57*** (s.e. 0.08–0.13); ∆D F coefficients -0.31*** to -0.36*** (0.07–0.09); interaction S×∆D F negative and in some specs significant (e.g., -0.40** (0.18)); Obs 5,532; R-sq. 0.03–0.05.

### Magnitudes, heterogeneity, mechanisms, and policy implications
- Magnitudes and heterogeneity (summary)
  - Average productivity losses of4-10percent due to power disruptions (aggregate statement across specifications).
  - Firms with high intensities of electricity use suffered higher productivity losses than firms at the middle of the power use distribution depending on the measure of power disruption.
  - Non-linear productivity losses across TFP quantiles: firms at higher TFP quantiles experienced higher productivity losses than those at the median.
  - Average factory closures ranging from1.4-2months in response to power disruptions.
  - No relationship was found between plant shutdown decisions and productivity cohorts (productive firms facing major power disruptions are as likely to rely on strategic shutdown as less productive ones).
- Mechanisms and explanations for higher losses at upper quantiles
  - Productive firms exposed to power disruptions are likely to self-generate during severe disruptions; higher productivity losses might be due to high self-generation costs.
  - Firms with power-intensive processes might face large capital adjustment costs when switching to power-saving technology and therefore absorb productivity losses.
  - Indirect effects through local supply chain disruptions, especially for power-intensive intermediate goods, can lead to lower capacity utilization.
- Policy implications and recommendations
  - Frequent power disruptions impose significant welfare loss in developing countries and can have debilitating effects on nascent manufacturing sectors and industrial development.
  - Maintaining an efficient power supply ensures productive firms remain productive and prevents locking up resources in unproductive activities.
  - Efficient power supply limits involuntary coping mechanisms such as variable capacity utilization or temporary suspension of operations to minimize losses.
  - Country-level analyses drawing on similarities in Sub-Saharan Africa can provide useful approximations in the absence of comparable cross-country firm-level panel data and can inform future meta-analyses on the subject.

*Source: wp18159*

### 3.1    Conceptual Framework

### 3.1    Conceptual Framework

### Power sector structure and firm choice environment
- Power generation and distribution in Ethiopia is entrusted to the state utility company; customers do not have the option to choose alternative power suppliers.
- Absent retail choice, firms facing power disruptions adopt coping mechanisms or absorb the full cost of outages.
- The conceptual decision tree (Figure 1) summarizes firm choices and implications for productivity:
  - Are electricity disruptions a major constraint on the firm’s operations?
    - No → Business as usual.
    - Yes → Coping mechanisms or shutdown.
      - Coping mechanisms:
        - Self-generate → ties up potentially productive capital K; may provide temporary relief but can reduce long-run output; self-generation does not guarantee complete substitution between grid and self-generated power.
        - Reduce capacity utilization (idle capacity) → small productive K; high cost of self-generation.
      - Shutdown → temporary cessation of production (idle capacity).

### Economic implications of self-generation and heterogenous firm responses
- Self-generation:
  - Allows replacement of grid electricity and continued production throughout the year for firms that can acquire generators.
  - Installation of self-generation capacity ties up potentially productive capital K.
  - The capacity to self-generate does not automatically guarantee complete substitution between grid and self-generated power.
  - The economic burden of self-generation may not be wholly offloaded to consumers through mark-ups if:
    - Firms face elastic demand for their goods.
    - Industry-specific tariff rates are low enough to allow cheaper imports.
  - Self-generation increases capital and operating costs, raises production costs, lowers competitiveness of local products, and limits potential for exploiting economies of scale (Steinbuks and Foster, 2010).
- Heterogeneity by firm size:
  - Self-generation option is often open only to medium and large firms.
  - Majority of smaller firms either:
    - Have small-scale operations making it too costly to own generators, or
    - Are rationed out of credit markets that facilitate such investments.
  - Small firms facing unanticipated disruptions are forced to scale down: lower intermediate input demand, vary capacity utilization over time, or undertake temporary shutdown.

### Measurement of power disruptions used in analysis
- Two alternative measures of power disruption:
  - S: dummy equal to one if the firm faced electricity disruption in a fiscal year (self-reported). Distinguishes ‘shortage-firms’ (firms ranking the problem as a major constraint) from ‘non-shortage firms’. The designation is time-varying.
    - Cautions:
      - Self-reported measures can underestimate the proportion of firms experiencing disruptions if generator-owning firms are less likely to report disruptions.
      - Timing of judgments about major constraints can induce endogeneity if firms decide on constraints after observing their productivities.
  - Sσ: dummy capturing all firms facing grid-power consumption drops exceeding ‘1standard deviation’ from long-run grid power consumption in any fiscal year (includes firms that cope or incur full cost).
    - Data shows Sσ is not correlated with other distortions that might lower grid-power consumption such as raw material and machinery-part shortages (see Appendix Table A.3).
- Observed relationship between measures:
  - The share of firms reporting power disruptions as a major constraint grew from below15percent of LMMI during2000−2006 to nearly25percent in2008 and33 percent in 2009.
  - The share of firms with electricity consumption (logs) below their long-run levels also increased.
  - The two measures track one another towards the end of the sample but are generally weakly correlated.
  - Except for 2003, Sσ > S; the difference Sσ − S may approximate firms that cope with power disruptions.

### Conceptual link to productivity and shutdown analysis
- The framework emphasizes two long-term implications of power disruptions:
  - Creation and perpetuation of idle capacity.
  - Tying up resources in unproductive capital (self-generation).
- These channels feed into empirical analysis by:
  - Quantifying productivity cost of power disruptions across firms varying by power use and productivity levels.
  - Examining the link between recurring blackouts and firms’ decisions to temporarily shut down to avoid losses.
- Productivity losses computed in the study are a sum total of all conceivable costs.

*Source: wp18159 - 3.1    Conceptual Framework (wp18159 - 3.1    Conceptual Framework)*

### 5.1    How costly are power disruptions?

### 5.1    How costly are power disruptions?

### Labor productivity losses — benchmark and robustness
- Pooled OLS benchmark (Table 3) finds:
  - Factor elasticities: capital 0.10***, raw materials 0.63***, electricity 0.07*** (variables in log-labor units; robust s.e. in parentheses).
  - Power disruption dummy: shortage-firms have "14-18percent" lower labor productivity than non-shortage firms (Columns 2 and 3).
  - Pooled OLS average estimates of productivity loss: "17-18percent" (Table 3 OLS averages).
- Fixed effects (FE) and IV comparisons (Table 4 and Appendix):
  - FE estimates of elasticities differ from OLS; FE β̃k (capital) is much lower than OLS.
  - IV (lagged investment per labor as instrument for capital) yields higher implied losses from power disruption (Appendix A.4 and Figure A.1); a standard Hausman test favors IV over OLS (FE), but IV suffers loss of information due to many zero-investment observations.
  - Power disruption dummies are negative but smaller under FE/IV: pooled OLS ~17-18percent vs FE/IV average productivity loss about "5-6percent" (columns 2 and 3 compared to OLS).

### Non-linear relationship with electricity intensity
- Extended specifications include electricity intensity ̃e and ̃e^2 and interactions with disruption dummy:
  - Column 4 of Table 3 and Table 4 include electricity^2 and interactions; squared term is marginally significant in some specifications, indicating non-linearity.
  - Predicted productivity losses plotted in Figure 3 (Table 4 Column 4 marginal effects):
    - Computation of losses: 1−exp(̃y_j^S − ̃y_j), where ̃y_j^S and ̃y_j are productivity with and without disruption at intensity j.
    - FE estimates yield average productivity losses of:
      - "6-7percent" at mean intensity,
      - "22-30percent" at low intensity,
      - "16-22percent" at high intensity.
  - Interpretation of inverted U-shaped loss curve (when S is disruption measure):
    - Firms at low intensity may find self-generation unaffordable and resort to strategic shutdowns.
    - Large/power-intensive firms may self-generate but incur high self-generation costs, producing higher losses.
  - Alternative disruption measure S_σ produces a different (downward sloping) loss curve (Appendix Table A.5 and Figure 4), with firms at higher intensities facing sharper losses under that measure.

### Total factor productivity (TFP) losses — control-function approaches and quantile regressions
- TFP estimation approaches:
  - Levinshohn and Petrin (LP) and Ackerberg et al. (ACF) control-function approaches used to back out TFP (Appendix Tables A.7−A.11).
  - Wide variance in estimated elasticities between LP and ACF when raw-material demand is the control function: LP labor and capital elasticities are about half the values of ACF in the raw-material control case.
  - Industry groups considered: chemical, glass and plastics (CGP); clothes and footwear (CAF); food and beverages (FAB); machinery, motors and accessories (MMA); pulp, paper and publishing (PPP).
- Panel quantile regressions at the median (Table 5):
  - MCMC with 50,000 draws per quantile used to estimate Eq.7.
  - LP-based TFP median results (Columns 1–3) indicate average productivity loss of "5-6percent" due to power disruptions.
  - ACF-based TFP median results (Columns 4–6) indicate average productivity loss of "nearly9percent" of TFP.
  - Table 5 coefficients (selected): Disruption coefficients reported as 0.06*, -0.05**, -0.05* (LP columns) and 0.02, -0.09***, -0.09*** (ACF columns); note: results at median, bootstrapped s.e. in parentheses.
- Within-industry (Table 6) median results:
  - First three point estimates (LP within-industry) indicate average productivity loss of "7percent" due to power disruptions.
  - ACF within-industry estimates show similar patterns though point estimates differ; disruption coefficients include -0.07*** and -0.03* in various columns.
- Productivity loss across quantiles (Figure 5):
  - Inverted U-shaped productivity loss curve across productivity distribution for aggregate TFP.
  - Firms at lower productivity quantiles Q_10 experience about "6 percent" loss; upper quantiles incur about "7-8percent" TFP losses.
  - Shape of loss curve varies with disruption measure S_σ (see Appendix Table A.14 and Figure A.5).

### Shutdown behavior as a coping response
- Empirical strategy: panel count-outcome models (Negative Binomial and Poisson) for months of shutdown I_it over firm age (Age as exposure variable).
- Descriptive patterns (Figure 6):
  - Plant shutdown episodes constitute "37percent" of the full sample (top-left panel).
  - Younger firms shutdown more often than mature firms.
- Count-model results (Table 7 and Appendix Table A.12):
  - Random effects NB (Columns 1–6) and conditional fixed effects NB (Columns 7–12) consistently show a positive and significant relationship between power disruptions and shutdown incidents.
  - On average, firms are likely to shutdown operations for "1.4-2months" in response to power disruptions.
  - Firms with history of unexploited capacity are more likely to shutdown:
    - Shutdown incidents increased by about "1.3months" when unexploited capacity in the past year was "33% higher".
  - Poisson alternative yields similar conclusions (Appendix Table A.12).
  - Productivity levels are not significant predictors of shutdown episodes across models and productivity concepts.
- Additional note:
  - Differenced panel data models and an alternative activity-period measure A = 12 − I are examined as further robustness checks (discussion continues beyond provided excerpt).

*Source: IMF Working Paper wp18159, Section 5.1*

### Appendix Table A.13 summarizes key results from a differenced OLS model. Again

### wp18159 - Appendix Table A.13 summarizes key results from a differenced OLS model. Again

### Model results: effect of power disruptions on activity months
- Results consistently show a negative and significant effect of power disruptions on activity months.
- Power disruptions are likely to compel firms to shutdown for nearly 1months.
- The results remain consistent when lagged second differences of productivity and excess capacity are used.
- A 33% increase in underutilized capacity is equivalent to1natural-log point increase inD f. From see Eq11, the response in months is exp(γ s ).

### Key negative binomial regression coefficients (selected)
- S: 0.61*** (0.05); 0.62*** (0.1); 0.67*** (0.11); 0.67*** (0.10); 0.69*** (0.10); 0.68*** (0.08); 0.57*** (0.05); 0.56*** (0.10); 0.63*** (0.13); 0.59*** (0.11); 0.60*** (0.10); 0.60*** (0.11)
- D F t−1: 0.27*** (0.05); 0.29*** (0.09); 0.23*** (0.06); 0.28*** (0.06); 0.23*** (0.06); 0.17*** (0.07); 0.13 (0.08); 0.14** (0.07); 0.15** (0.06); 0.15*** (0.06)
- ln TFP LP1 t−1: -0.1* (0.09); -0.11 (0.08)
- ln TFP ACF1 t−1: -0.05 (0.07); -0.01 (0.09)
- ln TFP LP2 t−1: 0.04* (0.08); -0.05 (0.09)
- ln TFP ACF2 t−1: 0.05 (0.09); 0.11 (0.09)
- Const.: -1.99*** (0.12); -2.61*** (0.17); -2.92*** (0.31); -2.74*** (0.19); -3.15*** (0.22); -2.90*** (0.19); -2.66*** (0.091); -3.10*** (0.16); -3.09*** (0.34); -3.26*** (0.22); -3.01*** (0.25); -3.70*** (0.20)
- Note: bootstrapped s.e. in parenthesis.

### Sample sizes, firms, and log-likelihood (by specification)
- Obs.: 11,094; 6,431; 4,856; 6,040; 6,040; 6,044; 6,418; 3,353; 2,151; 3,071; 3,071; 3,075
- Firms: 3,280; 1,927; 1,508; 1,789; 1,789; 1,790; 1,319; 778; 494; 689; 689; 690
- Log-Lik: -18320; -9112; -5829; -8140; -8136; -8144; -8149; -3945; -2506; -3539; -3539; -3540

### Variable definitions (as used in table)
- ln TFP LP(i) t−1 = lagged industry TFP (logs), i = 1 (proxy m)
- ln TFP ACF(i) t−1 = lagged industry TFP (logs), i = 2 (proxy e)
- D F t−1 = underutilized-capacity ‘ ln Y F − ln Y ’
- Estimates using the alternative power disruption dummy S σ are available upon request

### Productivity losses and shutdowns: magnitudes and heterogeneity
- Average productivity losses of4-10percent due to power disruptions.
- Firms with high intensities of electricity use suffered higher productivity losses than firms at the middle of the power use distribution depending on the measure of power disruption.
- Non-linear productivity losses across TFP quantiles: firms at higher TFP quantiles experienced higher productivity losses than those at the median.
- Average factory closures ranging from1.4-2months in response to power disruptions.
- No relationship was found between plant shutdown decisions and productivity cohorts (productive firms facing major power disruptions are as likely to rely on strategic shutdown as less productive ones).

### Mechanisms and explanations for higher losses at upper quantiles
- Productive firms exposed to power disruptions are likely to self-generate during severe disruptions; higher productivity losses might be due to high self-generation costs.
- Firms with power-intensive processes might face large capital adjustment costs when switching to power-saving technology and therefore absorb productivity losses.
- Indirect effects through local supply chain disruptions, especially for power-intensive intermediate goods, can lead to lower capacity utilization.

### Policy implications and recommendations
- Frequent power disruptions impose significant welfare loss in developing countries and can have debilitating effects on nascent manufacturing sectors and industrial development.
- Maintaining an efficient power supply ensures productive firms remain productive and prevents locking up resources in unproductive activities.
- Efficient power supply limits involuntary coping mechanisms such as variable capacity utilization or temporary suspension of operations to minimize losses.
- Country-level analyses drawing on similarities in Sub-Saharan Africa can provide useful approximations in the absence of comparable cross-country firm-level panel data and can inform future meta-analyses on the subject.

*Source: wp18159 - Appendix Table A.13 and accompanying text (excerpt).*

### chapter 4, pp. 235–308.

### wp18159 - chapter 4, pp. 235–308

### Data and context
- Source datasets and methods referenced in chapter: Enterprise Surveys (World Bank Group); Levinsohn-Petrin (LP) and Ackerberg-Caves-Fraser (ACF) estimators; fixed effects (FE) and instrumental variables (IV) approaches; balanced and unbalanced panel specifications; Poisson and pooled OLS specifications.
- Proxy variables used in production function estimation: Raw Mat. (raw materials) and Electricity as demand-side proxies in LP and ACF estimators.

### Unreliability of utility services to manufacturing firms (Table A.1)
- Firms experiencing electrical outages (%) 78.78 80.0
- Number of electrical outages in a typical month 8.50 8.20
- Average duration of a typical outage (Hours) 5.80 5.80
- Average losses due to electrical outages (% annual sales)†† 8.30 6.90
- Domestic shipping product loss (% total shipping) 2.00 1.20
- Percent of firms experiencing water insufficiencies 1.80 2.70
- Notes: †SSA average for latest available survey data between 2010-2017. ††Simple average across firms.

### Electricity cost shares by industry (Table A.2)
- Food & Beverage: 0.12 0.06 0.08 0.06 0.05 (2005–2009)
- Textile: 0.12 0.11 0.11 0.12 0.12
- Apparel: 0.02 0.02 0.16 0.18 0.18
- Leather Tanning & footwear: 0.02 0.03 0.03 0.03 0.04
- Wood Products: 0.24 0.23 0.18 0.09 0.17
- Paper & Paper Products: 0.05 0.06 0.05 0.05 0.06
- Chemicals: 0.04 0.04 0.03 0.03 0.03
- Rubber & Plastics: 0.09 0.06 0.06 0.04 0.06
- Glass & Non-metallic Minerals: 0.46 0.58 0.61 0.57 0.62
- Basic Iron, Steel, & Casting: 0.02 0.01 0.01 0.09 0.09
- Metal Products: 0.04 0.04 0.04 0.04 0.03
- Machinery & appliances: 0.01 0.01 0.01 0.01 0.02
- Motor vehicles & accessories: 0.03 0.03 0.01 0.02 0.03
- Furniture: 0.02 0.02 0.02 0.03 0.02
- Total: 0.11 0.10 0.11 0.10 0.12

### Correlations between shortage indicator Sσ and possible reasons for lower power consumption (Table A.3)
- Electricity — Electricity: 1
- Electricity — Raw material: 0.01
- Electricity — Machine parts: -0.03
- Electricity — Customers: 0.00
- Electricity — Working capital: 0.00
- Electricity — Machine breakage: -0.02
- Raw material — Raw material: 1
- Raw material — Machine parts: -0.12
- Raw material — Customers: -0.27
- Raw material — Working capital: -0.32
- Raw material — Machine breakage: -0.16
- Machine parts — Machine parts: 1
- Machine parts — Customers: -0.07
- Machine parts — Working capital: -0.09
- Machine parts — Machine breakage: -0.04
- Customers — Customers: 1
- Customers — Working capital: -0.20
- Customers — Machine breakage: -0.10
- Working capital — Working capital: 1
- Working capital — Machine breakage: -0.12
- Machine breakage — Machine breakage: 1

### Estimated productivity losses from electricity disruptions — selected results (Tables A.4, A.6, A.16, A.18)
- Unbalanced panel (Table A.4, IV Marginal effects: Col (8)): Productivity Loss at intensity levels (̃e):
  - Contrast (1 vs 0) 2: Productivity Loss -0.52
  - (1 vs 0) 3: Productivity Loss -0.42
  - (1 vs 0) 4: Productivity Loss -0.32
  - (1 vs 0) 5: Productivity Loss -0.23
  - (1 vs 0) 6: Productivity Loss -0.16
  - (1 vs 0) 7: Productivity Loss -0.13
  - (1 vs 0) 8: Productivity Loss -0.12
  - (1 vs 0) 9: Productivity Loss -0.15
  - (1 vs 0) 10: Productivity Loss -0.21
  - (1 vs 0) 11: Productivity Loss -0.29
  - Joint χ2 p val = 0.00 (Joint = 47.18)
- Unbalanced panel (Table A.6, IV Marginal effects: Col (6)): Productivity Loss at intensity levels (̃e):
  - (1 vs 0) 2: Productivity Loss 0.18
  - (1 vs 0) 3: Productivity Loss -0.02
  - (1 vs 0) 4: Productivity Loss -0.16
  - (1 vs 0) 5: Productivity Loss -0.25
  - (1 vs 0) 6: Productivity Loss -0.29
  - (1 vs 0) 7: Productivity Loss -0.31
  - (1 vs 0) 8: Productivity Loss -0.30
  - (1 vs 0) 9: Productivity Loss -0.26
  - (1 vs 0) 10: Productivity Loss -0.19
  - (1 vs 0) 11: Productivity Loss -0.07
  - Joint χ2 p val = 0.00 (Joint = 68.60)
- Balanced panel (Table A.16, IV Marginal effects: Col (8)): Productivity Loss at intensity levels (̃e):
  - (1 vs 0) 2: Productivity Loss -0.53
  - (1 vs 0) 3: Productivity Loss -0.37
  - (1 vs 0) 4: Productivity Loss -0.22
  - (1 vs 0) 5: Productivity Loss -0.10
  - (1 vs 0) 6: Productivity Loss -0.03
  - (1 vs 0) 7: Productivity Loss -0.04
  - (1 vs 0) 8: Productivity Loss -0.10
  - (1 vs 0) 9: Productivity Loss -0.23
  - (1 vs 0) 10: Productivity Loss -0.39
  - (1 vs 0) 11: Productivity Loss -0.54
  - Joint χ2 p val = 0.00 (Joint = 22.68)
- Balanced panel (Table A.18, IV Marginal effects: Col (8)): Productivity Loss at intensity levels (̃e):
  - (1 vs 0) 2: Productivity Loss 0.33
  - (1 vs 0) 3: Productivity Loss 0.14
  - (1 vs 0) 4: Productivity Loss 0.01
  - (1 vs 0) 5: Productivity Loss -0.09
  - (1 vs 0) 6: Productivity Loss -0.15
  - (1 vs 0) 7: Productivity Loss -0.19
  - (1 vs 0) 8: Productivity Loss -0.21
  - (1 vs 0) 9: Productivity Loss -0.21
  - (1 vs 0) 10: Productivity Loss -0.18
  - (1 vs 0) 11: Productivity Loss -0.14
  - Joint χ2 p val = 0.00 (Joint = 15.79)

### Fixed Effects and IV estimation highlights (selected coefficients)
- Unbalanced panel FE and IV estimates (Table A.5, Cols reported):
  - Capital: FE 0.06*** (p-values: (0.000)); IV 0.14** (0.019)
  - Raw Materials: FE 0.65*** (0.000); IV 0.60*** (0.000)
  - Water: FE 0.02*** (0.005); IV 0.00 (0.941)
  - Electricity: FE 0.07*** (0.000); IV 0.05** (0.017)
  - Disruption (Sσ): FE 0.01 (0.461); IV 0.67** (0.014) in some specifications
  - Disruption×Electricity: interaction negative and significant in some specs: -0.18** (0.052); -0.29** (0.011)
  - Observations range: 7,158; 6,994; 6,994; 2,464; 2,439; 2,439 across columns. Firms range: 2,360; 2,331; 2,331; 981; 976; 976.
  - R-sq. range: 0.72, 0.73, 0.73, 0.68, 0.68, 0.67
- Balanced panel FE and IV estimates (Table A.15 & A.17):
  - Capital: FE ~0.05**; IV ~0.06
  - Raw Mat.: FE ~0.62***; IV ~0.59–0.57***
  - Water: FE 0.02*–0.02**; IV ~0.01
  - Electricity: FE 0.07***; IV coefficients vary including 0.06** and 0.24** in some cols
  - Disruption and Disruption interactions show varied signs and significance across specifications.
  - Observations: 1,607 to 1,567 in FE columns; Firms: 212 to 191; R-sq. ~0.81–0.83.

### Aggregate and industry-level production function estimates (Tables A.7–A.11)
- Aggregate production functions (Table A.7):
  - LP estimator (proxy = Raw Mat.): l 0.29*** (0.01), k 0.16*** (0.02), m 0.60*** (0.01)
  - ACF correction (proxy = Raw Mat.): l 0.67*** (0.02), k 0.30*** (0.01), m 0.64*** (0.01)
  - LP/ACF with Electricity proxy: e 0.17*** (0.02) and e 0.22*** (0.02) reported in two columns.
  - Obs 8,372; Firms 2,684.
- Industry-level LP estimates (Table A.8) — coefficients (standard errors) for l, k, e:
  - CGP (Chemicals, Glass and Plastics): l 0.21*** (0.03), k 0.10*** (0.04), e 0.18*** (0.04); Obs 1,757; Firms 619
  - CAF (Cloth and Footwear): l 0.24*** (0.02), k 0.10** (0.04), e 0.15*** (0.04); Obs 1,069; Firms 273
  - FAB (Food and Beverages): l 0.34*** (0.03), k 0.14*** (0.03), e 0.16*** (0.03); Obs 2,386; Firms 728
  - MMA (Machinery, Motors & Accessories): l 0.27*** (0.02), k 0.24*** (0.03), e 0.20*** (0.04); Obs 2,308; Firms 861
  - PPP (Pulp, paper and printing products): l 0.34*** (0.05), k 0.17*** (0.06), e 0.19*** (0.05); Obs 843; Firms 201
- Industry-level ACF estimates (Table A.9) — larger labor and capital elasticities:
  - Examples: CGP l 0.63*** (0.06), k 0.43*** (0.07), e 0.28*** (0.04)
  - FAB l 0.69*** (0.08), k 0.30*** (0.03), e 0.26*** (0.04)
  - Obs and firm counts match Table A.8.

### Within-industry productivity loss and quantile heterogeneity (Tables/Figures A.2, A.3, A.4, A.5, A.3, Table A.14)
- Panel quantile regressions (Table A.14) using ACF proxy show:
  - Aggregate lnTFP A: Disruption coefficient 0.18* (0.02) in Col (1) and -0.05*** (0.01) in Col (2)
  - Within-industry lnTFP I: Disruption -0.04*** (0.01) in Cols (3)–(4)
  - Age×100 coefficient: -1.01** (0.00) in Col (1) and similar negative coefficients in industry specifications
  - Working Capital: 0.36** (0.02) in Col (1); smaller positive coefficients in other columns
  - Observations: 8,184; 9,943; 9,935; Firms: 2,649; 2,960.
- Figures A.2, A.3, A.4, A.5 (described in notes):
  - Predictive margins and contrasts plot firm predicted log labor productivity with and without power disruptions across INTENSITY = ̃e (x-axis).
  - Within-industry panels show TFP change (%) across quantiles Q10, Q25, Q40, Q50, Q65, Q70, Q85, Q90 with negative shifts for many quantiles when disruption dummy S or alternative Sσ is positive.

### Count and alternative models (Tables A.12, A.13)
- Poisson/random effects and conditional fixed effects models (Table A.12):
  - Disruption coefficient consistently positive and significant in many count specifications: e.g., Random Effects (1) Disruption 0.38*** (0.04) with Obs 11,094; Firms 3,280; Log Lik -19331.
  - Lagged underutilized capacity (D F t−1) often positive and significant: e.g., 0.17*** (0.04) in Col (1).
  - Lagged industry TFP terms (lnTFP LP and lnTFP ACF at lags) generally small and statistically mixed across specifications.
- Pooled OLS alternative (Table A.13):
  - Disruption: coefficients -0.46*** to -0.57*** across specifications (standard errors 0.08–0.13).
  - ∆D F (change in underutilized capacity): -0.31*** to -0.36*** (0.07–0.09).
  - Interaction S×∆D F: negative and in some specs significant (e.g., -0.40** (0.18)).
  - Inclusion of first-differenced output per labor ∆y l and ∆ lnTFP terms attenuate or explain variation; Obs 5,532; R-sq. 0.03–0.05.

### Notes on inference and robustness
- Multiple specifications (FE, IV, balanced/unbalanced panels, LP vs ACF, industry-level estimates, Poisson/count models, pooled OLS) are reported to assess robustness of the effects of power disruptions on firm-level productivity and industry TFP.
- Joint tests reported for productivity loss contrasts indicate statistically significant joint effects in several specifications (e.g., Joint χ2 values and p vals reported in Tables A.4, A.6, A.16, A.18).

*Source: wp18159 - chapter 4, pp. 235–308.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2018/wp18159.pdf_