## wpiea2021110-print-pdf

## Source details

**Canonical URL:** [wpiea2021110-print-pdf](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021110-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021110-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021110-print-pdf.pdf.json)

---

### Empirical context, identification and data
- Quasi-experiment: arrival of submarine cables (SCs) in SSA after 2009 provides exogenous variation in internet penetration via two instrumental variables:
  - IV1 (extensive margin): being directly connected to a SC in the post-2009 period (Treatment * Post2009).
  - IV2 (intensive margin): total installed capacity of submarine cables (gbps per second).
- Key identification assumptions:
  - Absent SC arrival during 2009-12, internet penetration trends in coastal (treated) and landlocked (control) countries would have remained broadly unchanged.
  - Timing and capacity of SCs depend little on macroeconomic conditions in connected countries and affect macro outcomes only through internet penetration.
- Data sources:
  - Internet penetration: International Telecommunication Union.
  - SCs mapping and capacity: cablemap.info and official SC websites; maps generated with Mapbox and OpenStreetMap.
  - Macroeconomic and sectoral data: IMF, United Nations, World Bank, ILO, Penn World Tables 9.1.
- Sample coverage:
  - Baseline sample covers 33 SSA countries over 2005-2017 (largest baseline regression).
  - Dataset is an unbalanced panel; sectoral regressions often use fewer countries (as few as 11 for some productivity regressions).

### First-stage evidence and instrument relevance
- Change in internet penetration, 2008–2017:
  - Full-sample average internet penetration: 5.0 percent in 2008 and 24.9 percent in 2017.
  - 2017 penetration range: 1.3 percent to 62 percent.
  - 2017 average penetration: treatment countries 28.3 percent; control countries 18.2 percent.
  - Change in penetration, 2008–2017: treatment countries +22 percentage points; control countries +15.2 percentage points.
- Simple DiD (2009-17 vs 2002-08): internet penetration rose 4.4 percentage points more in treatment than control countries (Annex Table A4).
- First-stage regression highlights (Table A5):
  - Treatment*Post2009: 3.1707* (1.7412).
  - Submarine cable capacity (gbps per second): 0.0008*** (0.0002).
  - Instrument relevance tests: high F statistics; Montiel-Pflueger effective F-statistic reported (14.548 and 14.675 in specifications).
- Controls in first stage include lagged real per capita GDP (log), mean years of schooling, trade openness, institutional quality (CPIA), credit to private sector, among others.

### Main macroeconomic impacts (2SLS IV estimates)
- Growth (real per capita GDP growth):
  - IV1 (SC arrival after 2009): no statistically significant impact.
  - IV2 (cable capacity): each percentage point increase in internet penetration → 0.37 percentage point increase in real per capita GDP growth (Internet penetration coefficient: 0.3749** in Table 1).
  - Table 1 sample specifics: (1) Observations 395 / No. of countries 32 (IV1); (2) Observations 265 / No. of countries 21 (IV2). R-squared: (1) 0.1119; (2) 0.3174.
- Productivity:
  - Aggregate TFP: each percentage point increase in internet penetration leads to around 1.5 percent increase in TFP (Table 5; Table A7 shows IV2 TFP coefficient 0.5324*** in one specification).
  - Aggregate labor productivity: impact insignificant at aggregate level in some specifications.
  - Sector labor productivity (IV2, selected):
    - Utilities: +6.9 percent per percentage point increase in internet penetration.
    - Trade: +1.4 percent.
    - Transportation: +4.9 percent.
    - Agriculture: negative and statistically significant (potential measurement issues from informality).
  - Table A7 and Table 5 report sectoral coefficients using cable capacity IV (IV2); sector regressions use fewer countries (11).
- Output structure (value added shares):
  - IV1 (arrival): higher internet penetration significantly:
    - Increases services share of total value added: coefficient 0.1634* (Table 2, column 1).
    - Reduces industry share of total value added: coefficient -0.3233** (Table 2, column 3).
    - Insignificant effect on agriculture share: coefficient 0.0365 (Table 2, column 5).
  - IV2 estimates for these outcomes are often insignificant.
  - Services subsector gains driven by “transportation, storage and communication”.
- Employment structure:
  - No significant impact on level of total employment (Table 4).
  - Sectoral employment shares:
    - Services employment share increases by 0.3-0.4 percentage points for each percentage point increase in internet penetration.
    - Agriculture employment share decreases by about 0.4 percentage points per percentage point increase in internet penetration.
    - Industry employment share: no significant impact.
  - Gendered effects: shift to services employment is two and a half times larger for women than men (e.g., Services, Female: IV1 0.4944***; IV2 0.4962***).

### Interpretation across margins and sensitivity to IV choice
- IV sensitivity:
  - Extensive margin (IV1: being connected after 2009) matters more for output and employment composition effects.
  - Intensive margin (IV2: cable capacity) matters more for growth and productivity effects.
  - Different samples under IV1 and IV2 complicate direct comparison; further research needed.
- Possible reasons for large IV2 growth effect:
  - Non-saturation of internet penetration in SSA relative to advanced economies.
  - New SC technology with much higher data capacity arriving after 2009.
- Key point estimates reiterated:
  - Each percentage point increase in internet penetration → 0.37 percentage point increase in real per capita GDP growth (IV2).
  - Internet penetration increases TFP substantially (~1.5 percent per percentage point of penetration in reported results).

### Robustness checks and caveats
- Parallel trends: pre-treatment DiD estimate is insignificant for 2000-2008 when setting 2005 as cutoff (Annex Table A6), supporting parallel trends.
- Controls and external shocks:
  - Results robust when controlling for terms of trade, excluding heavily affected countries from the 2008-09 global financial crisis, and adding access to electricity (Annex Table A8).
  - Estimates broadly unchanged when dropping high-digitalized-and-high-income countries (Annex Table A9).
- Sensitivity to institutional controls and sample selection:
  - Using CPIA (baseline) vs WGI regulatory quality (IRQ) affects significance: under IRQ, impact on GDP growth and TFP becomes statistically insignificant though signs remain expected; possible explanation is higher correlation between IRQ and IVs reducing IV strength.
  - Important caveat on external validity: CPIA-country sample covers slightly over 75 percent of SSA GDP; omission of non-CPIA countries (e.g., South Africa) may affect generalizability.
- Sample end-year sensitivity:
  - IV2 growth coefficients remain statistically significant across sample end-years except for 2013; coefficient size declines as sample extends toward later years when new cable capacity growth slows (Table A10).
- Instrument exogeneity evidence:
  - No systematic relationship between average income and cable deployment; SCs deployed regionally, often internationally funded, and cable capacity driven by available technology at deployment time.

### Policy implications and open questions
- Policy implications:
  - Expand access to high-quality and reliable internet connectivity in SSA to boost growth, raise labor productivity, and foster diversification of production and employment.
  - Complementary investments recommended: foundational electricity infrastructure and digital literacy skills.
  - Digitalization reforms can induce compositional shifts in employment toward services, with particularly large gains for women.
- Open research questions:
  - Whether service jobs are filled by displaced agricultural workers or by reallocation across sectors.
  - Whether increased services employment reflects higher wage returns for digitally ready workers.
  - For women: whether higher services employment reflects new labor force entry or sectoral shifts due to flexibility in service jobs.

*Source: Excerpt from the PDF chapter "Introduction" and Annex tables, wpiea2021110-print-pdf.*

### References .............................................................................................................

### References (excerpt) — Introduction and Overview

### Digitalization and internet as productivity driver
- Digitalization includes use of the internet to collect, store, process, and exchange information.
- Higher internet penetration is viewed as a positive shock to the productivity of both labor and capital.
- The internet shapes production and service delivery and is critical for business continuity (example: enabling telework during the COVID-19 pandemic).

### Internet access in Sub-Saharan Africa (SSA)
- Internet access in SSA in 2017 ranged from as low as 1.3 percent of the population to 62 percent.
- Average internet access in advanced economies in 2017 was 68 percent.
- The arrival of submarine cables, most of which landed in coastal SSA countries in 2009-2012, dramatically reduced access cost and increased download speed, fostering an escalation in internet penetration.

### Empirical challenge and identification strategy
- Prior studies reviewed in Minges (2015) use cross-country regressions to estimate the relationship between broadband internet penetration and economic growth.
- Some studies address endogeneity (e.g., Czernich et al. 2011; Katz and Callorda 2018); others do not (e.g., Qiang et al. 2009; Zaballos and López-Rivas 2012).
- The paper argues that arrival of submarine cables in SSA provides exogenous variation to construct instrumental variables for internet penetration:
  - Differentiated impact in coastal (treatment group) and landlocked (control group) countries.
  - Weak relationship between cables' data traffic capacity and the income level of connected countries.
- Section III elaborates on features of control and treatment groups.

### Research contribution and comparison to related studies
- The paper applies a modified submarine cables experiment to a larger sample of countries and more years than Hjort and Poulsen (2019).
  - Hjort and Poulsen (2019) assess firm-level impacts of fast internet on employment and productivity.
  - This paper focuses on country-level macro impacts: economic growth, productivity, and sectoral output and employment shares.
- Cariolle (2018) examines impact of submarine cables on internet penetration; this paper goes further by using cables-based estimates of internet penetration to assess macroeconomic impacts.
- To the authors' knowledge, this is the first paper to explore the submarine cables experiment to assess country-level macroeconomic impacts of internet penetration in SSA.

### Main empirical findings (summary)
- Internet penetration has a large and significant effect on real per capita GDP growth and productivity at aggregate and sector levels.
- Sectoral composition:
  - Higher internet penetration is positively associated with the share of services in the economy.
  - Higher internet penetration is negatively associated with the share of industry.
  - Effect is insignificant for the share of agriculture.
- Employment effects:
  - No significant impact on industry employment (percent of total employment).
  - Positive association with services employment.
  - Negative association with agriculture employment.
  - The positive association with services employment is higher for females than for males.

### Policy relevance and caveats
- Findings support renewed policy attention to digitalization reforms in SSA, including expanding access to high-quality and reliable internet connectivity to boost growth, raise productivity, and foster diversification of production and employment.
- Relevance is emphasized for SSA economies concentrated in production of primary commodities subject to cyclical volatility.
- Limitations:
  - Results are sensitive to the specification of the instrumental variable and the underlying samples.
  - The natural experiment is observational and not randomized in a strict sense; claiming causality requires fully controlling for relevant cross-country differences, which is challenging with observational data in SSA.

*Source: Excerpt from the PDF chapter "Introduction" (References and Annex listed), wpiea2021110-print-pdf.*

### Section  V and  VI report the  results and  discuss  their robustness,  respectively. Section  VII

### wpiea2021110-print-pdf - Section  V and  VI report the  results and  discuss  their robustness,  respectively. Section  VII

### II. LITERATURE REVIEW — substantive findings on digitalization
- Definition: Digitalization refers to the widespread use of digital technologies, such as the internet and mobile phones, to collect, store, analyze, and exchange information (World Bank 2016; Brookings 2017).
- Mechanisms:
  - Within growth accounting, digitalization can affect economic growth through improved productivity by complementing and nurturing human labor and capital (Barro 1991; Oulton 2010; Akerman et al. 2015; Tisdell 2017).
  - Empirical studies often use broadband internet penetration as a proxy for digitalization.
- Empirical estimates (country-level studies cited):
  - Qiang et al. (2009): a 10 percent increase in fixed broadband penetration is associated with a 1.2 percentage point increase in average real per capita GDP growth during 1980-2006 for developed countries; effect increases to 1.4 percentage points for developing countries.
  - Czernich et al. (2011): a 10-percentage point increase in broadband penetration increases per capita GDP growth by 0.9-1.5 percentage points (OECD countries).
  - Zaballos and López-Rivas (2012): a 10 percent increase in broadband penetration linked to a 3.19 percent increase in per capita GDP (26 Latin American and Caribbean countries; non-linear model).
- Non-linearity and saturation:
  - Katz and Callorda (2018): digitalization returns on growth follow an inverted U-shape, suggesting a saturation point; mobile broadband may have reached diminishing returns in many countries while fixed broadband remains in a phase of increasing returns.
  - Need to reach a sufficiently high level of digital penetration to reap digital dividends (Roller and Weavers 2001; Koutroumpis 2009).
- Structural transformation and distributional effects:
  - Technology-driven changes affect relative prices and productivity across sectors, triggering cross-sector labor and output shifts (Mathess and Kunkel 2020).
  - Cross-sector shifts may also result from changes in aggregate income, input-output or sectoral linkages, international trade, supply for skilled workers, transaction costs, and institutional soundness.
  - Houngbonon and Liang (2018): after broadband adoption in France, overall unemployment unchanged; services became net job creator and manufacturing net job loser.
  - Gender effects: Klonner and Nolen (2008) (South Africa) — women mainly drive employment gains from mobile networks; Sovbetov (2018) (Turkey) — increases in e-commerce associated with significant female employment growth.
- Productivity and distributional gains:
  - Dieppe (2020): technology-driven improvements have economically meaningful and statistically significant effects on labor productivity growth (country level).
  - At firm/individual levels, gains favor innovation (Franklin et al. 2009), highly productive firms (Paunov and Rollo 2016), and skilled workers performing non-routine tasks (Akerman et al. 2015).
- Labor market composition effects:
  - Digital technologies can complement or substitute human labor; employment impacts decompose into job creation and destruction.
  - Akerman et al. (2015): broadband roll-out displaced unskilled workers and reduced their wages.
  - Dieppe (2020): in the short run, job substitution effect dominates for both advanced economies and emerging markets and developing economies.
  - Many studies find positive employment impacts in services where tasks are less easily automated (Acemoglu 1999; Autor and Dorn 2009; Goos et al. 2014).
- Overall synthesis:
  - Previous studies broadly confirm positive impacts of digitalization on growth and productivity.
  - Magnitudes differ by level of digitalization, sector, skill level, and gender.

### III. The arrival of submarine cables in the coast of Africa — context and empirical relevance
- Submarine cables (SCs): fiber cables under the sea connecting countries and continents; provide faster speed and lower cost relative to satellite networks.
- Key events and patterns:
  - Mid-January 2020 Bloomberg reported: “African Internet Slows After Undersea Cables Break” — referencing failures of WACS and SAT3/WASC affecting SSA countries.
  - Most SCs in SSA became operational during 2009-12, though some countries had earlier connections (South Africa, Nigeria, Benin, Cameroon).
- Cable capacity evolution:
  - Cables arriving in SSA after 2009 have dramatically higher capacity.
  - Average cable capacity increased from 165 billion bits per second in 2002-2008 to above 6000 billion bits per second in 2009-2018.
- Measurable impacts observed after SC arrival:
  - Prior to SCs, internet access in SSA was on average 10 times more expensive than in other regions.
  - Arrival of EASSy helped expand internet access for 20 coastal and landlocked countries and lower broadband costs by as much as 90 percent (International Finance Corporation 2019; World Bank 2018).
  - Internet penetration (percent of population using the internet) increased on average from 8.6 percent in 2002 to about 25 percent in 2017, with most increase occurring after 2009.
  - Countries directly connected to a SC (mostly coastal) experienced a faster increase in internet penetration after 2009 than indirectly connected (landlocked) countries.
- Deployment features relevant for identification:
  - SCs are deployed regionally due to very high fixed costs and economies of scale; cables rarely skip intermediate countries along routes (example: ACE connects geographically continuous countries on western SSA coast).
  - Most SC projects have sourced funding internationally (examples: EASSy and ACE received large funding from the World Bank).
  - Coastal countries typically connected earlier and directly; landlocked countries connected later and indirectly through coastal backbone networks.
  - By end-2018:
    - (i) almost all SSA countries were connected to at least one SC either directly (coastal) or indirectly (through coastal countries);
    - (ii) most coastal countries were connected to multiple cables;
    - (iii) there were 14 active cables and many more expected to come onstream.
- Empirical confirmation:
  - Positive impact of SC arrival on internet penetration and speed in SSA confirmed in Cariolle (2018) and Hjort and Poulsen (2019).

### IV. Empirical strategy and data — identification, instruments, and model specification
- Proxy for digitalization:
  - Internet penetration defined as the percent of the population with access to the internet in a country; chosen because the related time series is the most consistently available for SSA countries.
- Outcomes estimated:
  - Real per capita GDP growth, productivity (total factor and labor productivity), sector value added shares, and employment shares.
- Quasi-experiment exploited:
  - The increase in internet penetration caused by SC arrival after 2009 and the increase in SC capacity over time are treated as a quasi-experiment largely independent of macroeconomic conditions.
  - “Treated” countries: those that became directly connected to a SC starting in 2009 (typically coastal).
  - “Non-treated” (control) countries: those indirectly connected to a SC (typically landlocked). Exceptions noted: Rwanda and Uganda are landlocked but directly connected (Rwanda via Kenya completed in 2010; Uganda connected in late 2009 with SEACOM via Tanzania).
- Instrumental variables constructed:
  - Two IVs capture:
    - (i) the unprecedented arrival of SCs starting in 2009 (extensive margin — being directly connected after 2009);
    - (ii) the data capacity of the SCs measured in bits per second (intensive margin — cable capacity).
  - Terminology: “data capacity”, “cable capacity” and “SCs capacity” used interchangeably.
- Identification strategy assumptions:
  - Absent the SC arrival during 2009-12, difference in internet penetration rates between coastal and landlocked countries would have remained broadly unchanged.
  - Timing and capacity of SCs depend little on macroeconomic conditions in connected countries; SCs impact macroeconomic outcomes only through internet penetration.
  - Potential confounders (e.g., trade openness differences, 2008-09 global financial crisis, 2014-15 commodity price shock) are controlled for in regressions.
- Model specification — two-stage least squares (2SLS):
  - First-stage fuzzy difference-in-differences specification (Equation (1)):
    - Internet_{i,t} = α0 + α1 Treatment_i + α2 Post2009_t + α3 (Treatment * Post2009)_{i,t} + α4 X_{i,t} + ε_{i,t}
    - Treatment_i = 1 if country i is directly connected to a SC; Post2009_t = 1 if time is 2009 onwards; (Treatment*Post2009)_{i,t} is the interaction; X_{i,t} is a vector of control variables including lagged per capita income, public investment, trade openness, terms of trade, population density, institutional quality, human capital, credit to private sector, and fiscal position.
    - α3 gives the average DiD effect on internet penetration of being directly connected to a SC after 2009.
  - Alternative first-stage with cable capacity and fixed effects (Equation (1’)):
    - Internet_{i,t} = γ0 + γ1 Cable_{i,t} + ∅_t + ϑ_i + γ2 X_{i,t} + μ_{i,t}
    - Cable_{i,t} is the data capacity of SCs in country i at time t; controls for time-fixed effects (∅_t) and country-fixed effects (ϑ_i).
  - Second-stage (Equation (2)):
    - Macro_{i,t} = ρ0 + ρ1 Internet̂_{i,t} + ρ2 X_{i,t} + ω_{i,t}
    - Internet̂_{i,t} is the fitted internet penetration from first-stage; Macro_{i,t} are selected macroeconomic indicators (real GDP per capita growth, TFP and labor productivity, sector shares in value added and employment).
  - Estimation details:
    - All equations estimated with robust standard errors clustered at the country level.
    - Equation (1) removes country- and time-fixed effects via the DiD specification; Equation (1’) explicitly includes them.
    - The two IVs capture different dimensions (extensive vs intensive margins) and are applied to different samples; both are treated as important and not substitutes.
- Controls and concerns:
  - Control variables vary slightly depending on whether outcome is growth/productivity or structural change (sector shares). For structural change regressions some controls (fiscal balance, private sector credit, population density) are not included per reviewed literature.
  - Confounders explicitly considered include trade openness, global financial crisis timing, commodity price shocks, and other country-specific factors.

*Italic: wpiea2021110-print-pdf (canonical source URL provided by the pipeline).*

### Annex Table A1 provides details on the data used in the empirical analysis. The data on internet

### Annex  Table  A3  presents  summary  statistics  for  the  main  variables  of  interest. Overall,  the

### Annex: Summary statistics and results (Table A3 and related tables)

### Data and summary statistics
- Full-sample average internet penetration: 5.0 percent in 2008 and 24.9 percent in 2017.
- 2017 penetration range: 1.3 percent to 62 percent.
- 2017 average penetration: treatment countries 28.3 percent; control countries 18.2 percent.
- Change in penetration, 2008–2017: treatment countries +22 percentage points; control countries +15.2 percentage points.
- Average year-on-year per capita GDP growth, 2000-2017: 1.9 percent (full sample); treatment countries 2 percent; control countries 1.7 percent.
- 2017 sector composition (value added / employment): treatment countries — service share of total value added 25.8 percent and employment 41.5 percent; control countries — value added 22.4 percent and employment 29.2 percent.
- Sample coverage note: The CPIA database covers all SSA countries except South Africa, Botswana, Namibia, Mauritius, Equatorial Guinea, Eswatini, Gabon and Seychelles. The countries in the sample account for slightly over 75 percent of SSA’s 2017 nominal GDP in PPP terms.

### Instrument relevance and exogeneity
- First-stage DiD (simple two-period two-group): internet penetration rose 4.4 percentage points more in treatment than control countries after 2009 (Annex Table A4).
- DiD in regression framework (Annex Table A5, column 1): DiD estimate = 3.2 (statistically significant).
- Treatment dummy alone is insignificant; becomes significant only when interacted with Post2009 (DiD).
- Within coastal-country sample, response of internet penetration to installed cable capacity: estimated coefficient = 0.0008 (positive and highly statistically significant; Annex Table A5, column 2).
- Control variables with significant and expected signs: real per capita GDP, years of schooling, institutional quality, credit to the private sector (other controls often insignificant).
- Instrument relevance tests: both arrival of SCs and cable data capacity pass with high F statistics (standard F and Montiel-Pflueger robust F).
- Instrument exogeneity evidence: no systematic relationship between average income and cable deployment (Figures 7 and 8); three reasons cited: (i) SCs deployed regionally (rare to skip a country), (ii) SCs usually internationally funded, (iii) cable capacity driven by available technology at deployment time.

### Macroeconomic impact

Growth effects
- IV results differ by instrument:
  - Using arrival of SCs after 2009 (IV1): no statistically significant impact of internet penetration on real per capita GDP growth.
  - Using total installed capacity of submarine cables (IV2): each percentage point increase in internet penetration leads to a 0.37 percentage point increase in real per capita GDP growth (Table 1).
- Table 1 key coefficients (dependent variable: Real per capita GDP growth):
  - Internet penetration coefficient: (1: IV1) 0.0530; (2: IV2) 0.3749** (standard errors in table).
  - Lagged real per capita GDP (log): (1) -0.8094; (2) -23.6526***.
  - Investment (% of GDP): (1) 0.0655***; (2) 0.0484.
  - Trade openness (% of GDP): (1) 0.0215*; (2) 0.1212***.
  - Credit to private sector (% of GDP): (1) -0.0716**; (2) -0.1920.
  - Fiscal balance (% of GDP): (1) 0.0967***; (2) 0.0823*.
  - Sample periods: 2005-17 for both specifications.
  - Observations / No. of countries: (1) 395 / 32; (2) 265 / 21.
  - R-squared: (1) 0.1119; (2) 0.3174.
- Interpretation note: the large IV2 growth effect may reflect non-saturation of internet penetration in SSA relative to advanced economies.

Output structure effects
- Using arrival of SCs (IV1): higher internet penetration significantly:
  - Increases services share of total value added (column 1, Table 2): coefficient 0.1634*.
  - Reduces industry share of total value added (column 3, Table 2): coefficient -0.3233**.
  - Has insignificant impact on agriculture share (column 5, Table 2): coefficient 0.0365.
- IV2 estimates for these outcomes are often insignificant.
- Subsector detail (Table 3): positive impact on services driven by “transportation, storage and communication” (transport/communication); industry subsectors (“mining and utilities”, “manufacturing”, “construction”) show negative signs consistent with aggregate industry decline but are individually statistically insignificant.
- Interpretation: internet penetration may raise productivity and lower relative prices in some services, increasing services’ share and reducing other sectors’ shares (share—not level—of industry value added declines).

Employment structure effects
- No significant impact on level of total employment (Table 4, columns 1 and 2).
- Sectoral employment shares:
  - No significant impact on industry employment share (columns 3 and 4).
  - Significant positive impact on services employment share and significant negative impact on agriculture employment share.
  - Estimated magnitudes: each percentage point increase in internet penetration leads to an increase of 0.3-0.4 percentage points in the share of services employment and a reduction of 0.4 percentage points in the share of agriculture employment (Table 4).
- Gendered effects within services:
  - Impact higher for females than for males: shift to services employment is two and half times larger for women than men.
  - Table 4 service employment coefficients (examples): Services: (7: IV1) 0.3377**; (8: IV2) 0.3627**. Services, Female: (11: IV1) 0.4944***; (12: IV2) 0.4962***.
- Interpretation: decline in agriculture employment share is broadly offset by increase in services employment share; the gendered pattern suggests larger gains in services employment for women.

Productivity effects
- Aggregate total factor productivity (TFP): each percentage point increase in internet penetration leads to around 1.5 percent increase in TFP (Table 5, column 1).
- Aggregate labor productivity: impact insignificant at aggregate level (Table 5, column 2).
- Sector labor productivity (Table 5, using IV2):
  - Utilities: labor productivity increases by 6.9 percent per percentage point increase in internet penetration (column 6).
  - Trade: labor productivity increases by 1.4 percent (column 8).
  - Transportation: labor productivity increases by 4.9 percent (column 9).
  - Agriculture productivity coefficient: negative and statistically significant (column 3) — potentially reflecting measurement error from informality in agriculture.
- Table 5 reported coefficients (selected): Individuals using the Internet (% of population) — Agriculture (1) 0.0147***; Mining (2) -0.0030; Manufacturing (3) -0.0212***; Utility (4) 0.0206; Trade (6) 0.0687**; Transportation (7) -0.0057; Finance (8) 0.0143**; Other services (9) 0.0493**.
- Note: Table 5 results are reported using cable capacity IV (IV2); results are not significant when using submarine cable arrival IV (IV1) or when specifying dependent variables in growth rates. Sector regressions use fewer countries (11) due to data availability.

Interpretation across margins
- Results are sensitive to the IV used:
  - Extensive margin (arrival of SCs / IV1) appears to matter more for output and employment composition effects.
  - Intensive margin (cable capacity / IV2) appears to matter more for growth and productivity.
- Caveat: different samples under the two IVs complicate direct interpretation; further research warranted.

### Robustness checks
- Parallel trends: pre-treatment DiD estimate is insignificant for 2000-2008 when setting 2005 as cutoff (Annex Table A6), supporting the parallel trend assumption.
- External shocks:
  - 2014-15 terms of trade shock: addressed by including terms of trade as a control variable.
  - 2008-09 global financial crisis: addressed by excluding heavily affected countries (South Africa, Nigeria, Mauritius, Angola and Kenya) in alternative regressions.
- Additional robustness:
  - Estimates broadly unchanged when controlling for access to electricity.
  - Estimates robust to dropping high-digitalized-and-high-income countries.
  - Estimates robust to restricting sample endpoint to earlier years.
  - Some regressions sensitive to alternative controls for institutions.

*Source: Annex, wpiea2021110-print-pdf*

### Annex Table  A7 shows  the

### wpiea2021110-print-pdf - Annex Table  A7 shows  the

### Robustness of main results
- The estimated coefficient on internet penetration remains sizeable and significant for the growth regression. 27
- The coefficient becomes insignificant for the productivity regression in one specification; this may reflect the much lower sample size compared to the already small initial size.
- Results for the remaining value added and employment regressions (not shown in the paper) also do not change significantly. 27

### Access to electricity
- Digitalization is intrinsically related to electricity given that digital equipment typically requires electrical power to function.
- Concern: countries with higher submarine cables (SCs) capacity and greater availability of other digital equipment may also have higher access to electricity, and electricity as a direct input to production could bias regressions if omitted.
- The paper argued earlier that the arrival of the SCs and their capacity are exogenous to countries’ income status, implying in principle exogeneity to electricity access as well.
- Annex Table A8 shows growth and productivity regressions with access to electricity added as a control variable; the estimated coefficient of internet penetration remains broadly unchanged and statistically significant for both the growth and productivity regressions.

### Income-driven digitalization (reverse causality)
- Concern: fast-growing and higher-income countries may have adopted digital policies earlier, potentially upward-biasing the estimated impact of internet penetration on growth.
- Mitigating factors:
  - The IVs are argued to be exogenous to growth and income.
  - The cable arrival IV regression is set in a DiD scheme which removes fixed effects such as countries growing constantly faster over time.
  - The analysis directly controls for determinants of growth as in standard growth regressions.
- Annex Table A9: additional regressions exclude the “champions” of internet penetration (those with very high penetration rates above 40 percent in 2017, compared to the sample average of 25 percent): Botswana, Cabo Verde, Côte D’Ivoire, Gabon, Mauritius, Namibia, Nigeria, Senegal, Seychelles and South Africa. 28
  - The estimated coefficient of internet penetration on growth and productivity remains statistically significant and sizeable.

### Sensitivity to sample end-year (treatment window)
- Most treatment countries became connected between 2009 and 2012; the increase in total cable capacity also occurred during the same period.
- Identification takes the treatment period as the whole post-2008 period, assuming part of the increase in internet penetration during 2013-17 reflects earlier SC arrivals and capacity increases, consistent with observed lagged effects to technology shocks.
- Annex Table A10 reports growth regression coefficients using the cable capacity IV across different sample end-years:
  - The coefficients remain statistically significant irrespective of the sample end-year, except for the year 2013.
  - The size of the coefficients decline as the sample extends towards more recent years when the number of new cables and their capacity remained broadly unchanged. 29
- Similar findings for the productivity regression, except that the coefficient is also significant when 2013 is the sample end-year. 29

### Sample selection and institutional controls
- The control variable for institutional quality restricts the sample to CPIA countries. CPIA countries account for slightly over 75 percent of SSA GDP; omission of some countries like South Africa (19 percent of SSA GDP) could cause sample selection bias.
- Alternative: World Bank’s World Governance Indicators—index of regulatory quality (IRQ) has greater country and year coverage.
  - Under IRQ specifications, the impact of internet penetration on real per capita GDP growth and total factor productivity becomes statistically insignificant, albeit maintaining the expected sign and with a still significant IV in the first-stage regression (not reported).
  - Coefficients remain significant for labor productivity at the aggregate level and for selected sectors; manufacturing gains significance and transportation loses it (Annex Table A11).
  - Significance remains broadly unchanged for regressions on sector value added (Annex Table A12-13) and employment shares (Annex Table A14).
  - The higher correlation between the IRQ and the IVs (compared to the CPIA), which leads to less significant IVs, could explain the loss of significance in some second-stage regressions.
- Important caveat: external validity of baseline findings may not necessarily generalize to non-CPIA countries. 30

### Key empirical findings
- Each percentage point increase in internet penetration leads to a 0.37 percentage point increase in real per capita GDP growth, all else equal.
- The large estimate may reflect:
  - Sample consists entirely of SSA countries where returns to digital technologies are likely higher.
  - Recently arrived submarine cables in SSA use new technology with higher data capacity; quality effects may drive the larger estimate.
- Internet penetration increases labor productivity, particularly in the sectors of utilities, trade, and transportation.
- Internet penetration significantly boosts the share of services (as percent of total value added), with a positive impact on subsectors of “transportation, storage and communication”.
- Higher internet penetration reduces the share of industry; no significant impact on the share of agriculture in output.
- Internet penetration does not have a significant impact on the total level of employment.
  - It has a significant impact on services employment as a share of total employment: the services share increases by 0.3-0.4 percentage points for each percentage point increase in internet penetration.
  - The impact on services employment is higher for female than for male—the shift to services is two and a half times larger for women than men.
  - Agriculture and industry employment impacts are, respectively, negative and insignificant.

### Interpretation and open questions
- Findings imply higher internet penetration affects the composition, but not the level, of total employment:
  - Net gains in services sector jobs broadly offset agricultural job losses; impact on industry jobs is neutral.
- Important questions for follow-up research:
  - Whether new service jobs are filled by displaced agriculture workers or by industry workers moving into services and agriculture workers moving into industry.
  - Whether the increase in the share of services employment reflects higher wage returns in services driven by more digital-ready workers.
  - Within the female sample, whether the increase in services employment share is due to (i) new women joining the labor market or (ii) women moving from other sectors into services because of higher flexibility in service jobs allowing better reconciliation of household work.

### Instrument sensitivity and interpretation
- Results are sensitive to the instrumental variable used for internet penetration:
  - Estimated effects on GDP growth and productivity only hold when employing submarine cable capacity as the instrumental variable.
  - The two IVs both relate to submarine cables but measure different margins of adoption (extensive vs. intensive) and entail different samples.
  - One potential implication: the intensive margin (intensity of connectivity to submarine cables) may matter more to growth and productivity than merely being connected to the cables (extensive margin).
  - However, different sample composition under the two IVs challenges this interpretation; further in-depth research is warranted.

### Policy implications
- Findings support renewed attention to digitalization reforms in SSA to:
  - Expand access to high-quality and reliable internet connectivity as a tool to boost growth, raise labor productivity, and foster diversification of the production structure.
  - Digitalization is particularly relevant in SSA where economies are highly concentrated in primary commodities subject to cyclical volatility.
- Labor market implications:
  - Digitalization reforms could lead to compositional shifts in employment towards services for both men and women.
  - For SSA women particularly, higher internet connectivity could pave the way for more flexible work arrangements and more online business opportunities, potentially explaining why women largely drive the employment shift toward services.
- Complementary investments required:
  - Foundational infrastructure such as electricity.
  - Critical investments in digital literacy skills to ensure workers can take advantage of new digital opportunities.

*Source: wpiea2021110-print-pdf - Annex Table  A7 shows  the*

### References

### wpiea2021110-print-pdf - References

### Key empirical findings from tables
- Table A4 — Difference-in-Differences in Internet Penetration (percent):
  - (a) 2009-17: Treatment 16.4; Control 9.5; Difference 7.0
  - (b) 2002-2008: Treatment 4.1; Control 1.5; Difference 2.6
  - (a) - (b): Difference 12.4; 8.0; 4.4
- First-stage IV estimates (Table A5):
  - Treatment*Post2009: 3.1707* (1.7412)
  - Treatment: -1.5677 (1.3320)
  - Post2009: 3.6847*** (1.1392)
  - Submarine cable capacity (gbps per second): 0.0008*** (0.0002)
  - Lagged real per capita GDP (log): 4.7816*** (1.1640) and 15.3078** (7.4566) across specifications
  - Mean years of schooling: 0.9560** (0.4377) and 4.2246 (6.5026)
  - Trade openness (% of GDP): -0.0319 (0.0217) and -0.0917* (0.0506)
  - Institutional quality (CPIA rating): 2.8751** (1.1846) and 1.0618 (2.8658)
  - Credit to private sector (% of GDP): 0.2773*** (0.0777) and 0.3369 (0.3204)
  - Sample: 2005-17; Observations: 395 and 265; R-squared: 0.5920 and 0.8520; F-statistic: 12.79 and 2.95; Montiel-Pflueger effective F-statistic: 14.548 and 14.675; No. of countries: 32 and 21
- Internet penetration and growth / productivity (selected IV results across robustness tables):
  - Table A7 — Total factor productivity (log):
    - (1: IV1) 0.1033 (0.0707)
    - (2: IV2) 0.5324*** (0.1863)
    - (3: IV2) 0.0028 (0.0029)
    - Samples: 2005-17, 2005-17, 2005-18; Observations: 360, 230, 116; R-squared: 0.1306, 0.1790, 0.7058
  - Table A10 — Real per capita GDP growth (selected):
    - (1: IV1) 0.0396 (0.0651)
    - (2: IV2) 0.3685** (0.1655)
    - (3: IV2) 0.0174*** (0.0038)
    - Samples and settings vary: Sample 2005-17 and 2005-18; Observations: 394, 264, 151; R-squared: 0.1183, 0.3226, 0.4157
  - Table A10 — Total factor productivity (log) (additional):
    - (1: IV1) 0.0794 (0.0968)
    - (2: IV2) 0.6957** (0.3246)
    - (3: IV2) 0.0246* (0.0141)
    - Sample: 2005-17; Observations: 213, 213, 112; R-squared: 0.1152, 0.3317, 0.4665

### Data, sample coverage, and instruments
- Note on instruments (Tables A5, A6, A7, A8, A9, A10): IV1 = being directly connected to a submarine cable in post-2009 period; IV2 = total installed capacity of submarine cables (gbps per second).
- Table A2 — Country coverage, 2005-17: Lists 48 SSA countries with sample inclusion flags (“Yes”, “Yes*”, “No”), notes:
  - All “Yes” countries make up the sample under IV1.
  - For IV2 (cable capacity), the sample covers coastal countries directly connected to a submarine cable (“Yes*”), including Rwanda and Uganda.
  - Sectoral productivity regressions drop to 11 countries due to data limitations (Angola, Cameroon, Ghana, Kenya, Mozambique, Nigeria, Rwanda, Senegal, Sierra Leone, Tanzania, Uganda).
- Note: The vintage of most databases is as of October 2019.
- Main data sources (Annex): ITU, IMF, World Bank, ILO, Penn World Tables 9.1, Greg's Cable Map, and authors' calculations.

### Descriptive statistics (Table A3 highlights)
- Full sample (Mean, Mean 2008, Mean 2017, Std. Dev., N):
  - Internet use (percent of population): 8.6; 5.0; 24.9; 12.2; 789
  - Real GDP (Log): 7.9; 7.9; 8.0; 1.0; 836
  - Growth of real GDP per capita (%): 1.9; 2.1; 1.2; 4.9; 792
  - Share of agriculture of total value added (%): 23.8; 24.0; 21.4; 15.5; 792
  - Share of service in total value added (percent): 23.3; 23.5; 24.6; 8.0; 792
  - Share of manufacturing employment in total (%): 12.2; 11.9; 12.7; 6.9; 817
  - Share of service employment in total (%): 34.5; 33.7; 37.2; 15.9; 817
  - Total investment (percent of GDP): 22.8; 23.5; 22.3; 10.2; 808
  - Trade openness (percent of GDP): 75.5; 80.5; 72.8; 38.2; 770
  - Population density (people per square km): 97.8; 94.5; 116.8; 122.7; 829
  - Average years of schooling: 4.6; 4.6; 5.3; 2.1; 775
  - Credit to private sector (percent of GDP): 26.0; 19.4; 72.7; 114.8; 809
- Treatment vs Control subsamples (Table A3):
  - Treatment Internet use (percent): 10.2; 6.3; 28.3; 13.5; 521
  - Control Internet use (percent): 5.4; 2.5; 18.2; 8.3; 268
  - Treatment population density: 116.8; 113.7; 137.0; 134.0; 551
  - Control population density: 60.1; 57.3; 75.0; 85.2; 278

### Sectoral value-added and employment results (selected coefficients)
- Table A12 — Internet penetration and sector value added (coefficients on Individuals using the Internet (% of population)):
  - Agriculture: 0.0049* (0.0029)
  - Utility: 0.0314*** (0.0121)
  - Construction: 0.0410*** (0.0146)
  - Finance: 0.0133* (0.0077)
  - Other services: -0.0026 (0.0059)
  - Sample: 2002-17; Observations: 222 for each sector; No. of countries: 14; Year/Country fixed effects: Yes; R-squared varies by sector (report shows e.g., Agriculture 0.9892)
- Table A14 — Internet penetration and employment shares (selected coefficients on Individuals using the Internet (% of population)):
  - Trade, restaurants, hotels: 0.0726 (0.0561)
  - Transport, storage & communication: 0.1185*** (0.0306)
  - Mining and utilities: -0.3482*** (0.1343)
  - Manufacturing: -0.0163 (0.0543)
  - Construction: 0.0173 (0.0340)
  - Sample: 2002-17; Observations: 602 for many employment regressions; No. of countries: 40 (or 26 in IV2 specifications); Year FE: sometimes No 1/ (difference-in-differences first stage removes fixed effects)
- Labor productivity and labor productivity by sector (Table A12 and continuation):
  - Labor productivity (log), coefficient on Individuals using the Internet (% of population): 0.1912*** (0.0627) in one specification; other sectoral labor productivity coefficients include negative and positive significant estimates across industry, agriculture, services (varies by IV and fixed-effects inclusion).
  - Many regressions include controls: lagged real per capita GDP (log), investment (% of GDP), mean years of schooling, trade openness (% of GDP), regulatory quality, terms of trade (change), population density, credit to private sector (% of GDP), fiscal balance (% of GDP).

### Robustness, placebo, and sensitivity tests
- Placebo test in 2005 (Table A6): placebo specifications across samples (2000-2017; 2000-08) and alternative institution controls (WGI regulatory quality) produce consistent results (detailed coefficient tables in source).
- Table A7: Results reported removing countries heavily affected by the 2008 global financial crisis (South Africa, Mauritius not in baseline; other heavily impacted: Nigeria, Angola, Kenya).
- Table A8: Regressions controlling for access to electricity as additional control; IV definitions and difference-in-differences first-stage applied.
- Table A9: Sensitivity excluding high internet penetration countries (Botswana, Mauritius, Namibia, South Africa, Gabon, Seychelles, Cabo Verde, Cote d’Ivoire, Nigeria, Senegal).
- Table A10: Sensitivity on sample end-year shows sample end-year coefficients for Internet use (% pop.) by sample end-year: 
  - Sample end-year 2012: 0.9608* (0.5701)
  - 2013: 0.9709 (0.6560)
  - 2014: 0.8355* (0.4324)
  - 2015: 0.6782** (0.2766)
  - 2016: 0.5018** (0.2185)
  - 2017: 0.3749** (0.1661)
  - Observations by sample: 166, 187, 207, 226, 245, 265; R-squared reported per sample.

### Notes on significance and methodology
- Statistical significance notation used throughout: ***, ** and * denote statistical significance at 1%, 5% and 10% level, respectively.
- Standard errors are in parenthesis and are clustered at the country level.
- Many IV1 specifications use a two-period two-group difference-in-differences specification in the first stage that removes fixed effects (noted as No 1/ in several tables).
- Year/Country fixed effects are included in many, but not all, specifications; tables indicate which specifications include fixed effects.

*Source: References and Annex tables from wpiea2021110-print-pdf*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021110-print-pdf.pdf_