## Annex I. Potential Determinants of Digital Adoption

## Source details

**Canonical URL:** [Annex I. Potential Determinants of Digital Adoption](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023065-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023065-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023065-print-pdf.pdf.json)

---

### Variables and measurement
- Data elements (Name — Source — Unit):
  - ICT Investment — World Bank — % GDP
  - Population Density — World Bank — 100 per sq km
  - Electricity Access — World Bank — %
  - Log Fixed Broadband — ITU — $ PPP
  - Log Mobile High Package — ITU — $ PPP
  - Log Mobile Low Package — ITU — $ PPP
  - Industry Value Added — World Bank — % GDP
  - Services Value Added — World Bank — % GDP
  - Agriculture Value Added — World Bank — % GDP
  - GDP per Capita — World Bank — $1,000
  - Education Completion (Secondary) — World Bank — %
  - Education Completion (Primary) — World Bank — %
  - Women Empowerment — World Bank — %
  - Inequality — World Bank — GINI
  - Poverty at $1.9 — World Bank — Head-Count Ratio ($1.9)
  - Rural Population — World Bank — %
  - Lowest 20% Income — World Bank — %
  - 2nd Lowest 20% Income — World Bank — %
  - 3rd Lowest 20% Income — World Bank — %
  - 2nd Highest 20% Income — World Bank — %
  - Highest 20% Income — World Bank — %
  - Labor Force Participation Rate — World Bank — %
  - Population Covered 3G (%) — ITU — %
  - Handset Price — GSMA — % Monthly GDP
  - Gender Parity — GSMA — %
  - UN E-Government Score — UN — Score (1-100)
  - Apps in National Language(s) — GSMA — %
  - Mobile Ownership — GSMA — % Population

### Data interpolation, extrapolation, and imputation
- Study period covered: 1997-2021.
- Missing WDI handling:
  - Linearly interpolate and extrapolate using the gradient from the two most recent years available for each variable.
  - Negative extrapolated values replaced with missing values.
  - For variables measured in percentages, extrapolated values above 90 percent replaced with missing values.
- To enable multivariate regressions despite remaining missing values:
  - Replace missing values with average values (means).
  - Generate a dummy variable that equals one for observations imputed using means.
  - Include these dummy variables as controls in regressions.

### Correlates of internet adoption (summarized)
- Dependent measures used in figures:
  - Percent of population using internet; percent of population with internet access at home.
- Estimation notes:
  - Regression coefficients β1 reported from specified equations.
  - Controls include GDP per capita.
  - Fixed effects: Income Group x Year fixed effects or Country fixed effects; standard errors clustered at country level.

### Major empirical findings on digital adoption (cross-country patterns)
- Digital adoption measure: share of population who report using the Internet in past three months.
- Cross-country disparities (end-2020):
  - AEs: 90 percent Internet use.
  - EMEs: 72 percent Internet use.
  - LIDCs: 32 percent Internet use.
  - Fixed broadband subscriptions: AEs 36 percent, EMEs 14.5 percent, LIDCs 2 percent.
  - Mobile phone ownership: AEs 95 percent, EMEs 80 percent, LIDCs 56 percent.
- Coverage and price patterns:
  - 3G signal coverage and cell phone connectivity: EMEs and LIDCs are close to AEs.
  - 4G connectivity: large gaps, especially between AEs and LIDCs.
  - Effective prices of using the internet (PPP-adjusted): highest in LIDCs despite recent declines.
- Conditional correlates of Internet use (controlling for GDP per capita, within-country variation):
  - Positive correlations: ICT Investment; Coverage of 3G; Mobile phone ownership; Services share of GDP; Secondary education completion; Apps in national languages; Labor force participation; Urbanization; Share of income held by poorest 20 percent; UN E-Gov Score; Gender equality (women empowerment).
  - Negative correlations: Variable costs (price of broadband and mobile data); Poverty and inequality; Rural population; Agriculture and Industry value added.
- Sample and data: 4,124 country-year observations, 169 countries, 1997–2021 (interpolated/extrapolated within a reasonable range).

### Machine-learning selection and policy classification
- LASSO-selected strongest predictors (conditional on GDP per capita and country fixed effects) preserve directions from multivariate regressions.
- Two policy-relevant groups:
  - Fast-moving variables (short-run policy levers): broadband and mobile data prices; smart phone ownership (mobile ownership); coverage; use-cases/apps in national languages.
  - Slow-moving variables (require structural policy): skills (secondary education completion), labor force participation, inequality, urbanization, services share of GDP.
- Policy implication: fiscal policy can target fast-moving cost variables (e.g., subsidies to lower connection cost) while recognizing slow-moving complements that require structural reforms.

### Model of demand for internet use and price responsiveness
- Consumer utility specification: logit framework incorporating country fixed effects, year fixed effects, market demand shock, and income-group-specific price sensitivities αb for b ∈ {L (LIDC), E (EME), A (AE)}.
- Endogeneity approach: control-function using leave-one-out regional price averages P−ct as supply shifter; robustness checks use lagged IVs.
- Implied average demand elasticities (reported summary):
  - AE: -0.004
  - EME: -0.26
  - LIDC: -1.22
- Interpretation: elasticity (absolute value) largest in LIDCs, moderate in EMEs, nearly price-inelastic in AEs; demand elasticity decreases with income level and over time.

### Fiscal costs of subsidies to increase internet use (counterfactuals)
- Counterfactual targets:
  - Raise Internet use in LIDCs to the level of EMEs.
  - Raise Internet use in EMEs to the level of AEs.
- Cost estimates assuming subsidies are untargeted:
  - EMEs: 0.3 percent of annual GDP.
  - LIDCs: 3.2 percent of annual GDP.
- Cost estimates assuming subsidies are targeted to those deprived of internet:
  - EMEs: 0.1 percent of annual GDP.
  - LIDCs: 1.8 percent of annual GDP.
- Graphical intuition: untargeted subsidies imply larger public expense (shaded area under demand curve); perfect targeting reduces public expense but may involve administrative costs and residual deadweight loss.

### Section IV targeting guidance — price-discrimination analogy and cost calculations
- Untargeted subsidies analogous to no price discrimination (one price for all).
- Targeted subsidies analogous to third-degree price discrimination (subsidized price for non-adopters).
- Perfect targeting analogous to first-degree price discrimination (government fills gap between willingness to pay and actual price).
- Cost estimates and examples:
  - To increase internet use for EMEs from 72 to 90 percent requires a 96 percent subsidy (18/72/0.26).
  - To increase internet use for LIDCs from 32 to 70 percent requires a 97 percent subsidy (38/32/1.22).
  - If untargeted, these subsidies amount to:
    - 0.34 percent of annual GDP for EMEs.
    - 3.2 percent of annual GDP for LIDCs.
  - If targeted, corresponding amounts decline to:
    - 0.07 percent of annual GDP for EMEs.
    - 1¾ percent of annual GDP for LIDCs.
- Assumptions used:
  - Average monthly price of mobile internet: $4 in EMEs and $16 in LIDCs (PPP, ITU).
  - Population numbers (averages over income group, year 2020): 62.5 million in EMEs and 29.7 million in LIDCs.
  - GDP PPP in constant 2017 dollars: $122 billion for an average LIDC and $827 billion for an average EME.
  - Subsidy calculation: number of people given subsidies × subsidy percentage × 12 × monthly cost of mobile internet.

### Market structure and supply-side responses
- Regression of price of mobile high basket on number of operators (controlling for country income-group fixed effects and log population) shows a negative but weak relationship between number of operators and mobile internet price.
- Implication: market concentration likely affects equilibrium prices; price subsidies may elicit supplier price responses, especially in highly concentrated markets.
- Policy stance: proceed conservatively due to possible supplier responses.

### Composition of the digital divide and targeting groups (FINDEX 2017 evidence)
- Data source: World Bank FINDEX survey, 2017, ~1,000 individuals per country, with sampling weights.
- Cross-country simple averages:
  - Mobile phone ownership and debit card ownership lower in LIDCs than EMEs and AEs.
  - LIDCs lag in internet-based activities: internet banking, online bill payments, e-commerce, ability to get emergency funds.
- Promising LIDC usages:
  - Sending and receiving domestic remittances highest in LIDCs; almost half via phones.
  - About 16 percent of government transfers in LIDCs are received via phones—higher than EMEs and AEs.
- Summary statistics (Table 4; 2017; Standard Deviations in parentheses):
  - Owns Mobile Phone: LIDC 0.660 (0.474); EME 0.847 (0.360); AE 0.935 (0.247).
  - Has Debit Card: LIDC 0.125 (0.331); EME 0.431 (0.495); AE 0.848 (0.359).
  - Used Phone/Internet for Banking: LIDC 0.062 (0.241); EME 0.181 (0.385); AE 0.588 (0.492).
  - Online Bill Payments: LIDC 0.054 (0.225); EME 0.150 (0.357); AE 0.559 (0.497).
  - Bought Online: LIDC 0.036 (0.186); EME 0.144 (0.351); AE 0.544 (0.498).
  - Received Government Transfers: LIDC 0.066 (0.248); EME 0.132 (0.338); AE 0.277 (0.448).
  - Sent Domestic Remittance: LIDC 0.226 (0.418); EME 0.145 (0.352); AE 0.042 (0.201).
  - Sent Remittance by Phone: LIDC 0.112 (0.315); EME 0.033 (0.179); AE 0.005 (0.070).
  - Observations: LIDC 426,437; EME 212,034; AE 56,740.

### Within-country determinants of mobile ownership (Table 5 highlights)
- Regression: Owns Mobileict = β0 + β1 Xict + γc + δt + εict (country and year fixed effects; cluster SEs at country level).
- Key coefficients (2017):
  - Female on mobile ownership:
    - LIDC: -0.090*** [0.005]
    - EME: -0.062*** [0.004]
    - AE: -0.062*** [0.006]
  - Secondary Education:
    - LIDC: 0.168*** [0.007]
    - EME: 0.083*** [0.004]
    - AE: 0.019*** [0.005]
  - Tertiary Education:
    - LIDC: 0.214*** [0.016]
    - EME: 0.091*** [0.006]
    - AE: 0.035*** [0.006]
  - Female × Secondary and Female × Tertiary: positive and significant across income groups.
  - In Workforce:
    - LIDC: 0.069*** [0.005]
    - EME: 0.058*** [0.003]
    - AE: 0.033*** [0.003]
  - Dependent variable means: LIDC 0.69; EME 0.85; AE 0.95.
  - R^2: LIDC 0.181; EME 0.161; AE 0.104.
- Interpretation:
  - Women less likely to own mobile phones across income groups; gap especially large in LIDCs.
  - Mobile ownership increases monotonically with education.
  - Older individuals (≥ 65) less likely to own mobile phones.
  - Richer individuals and workforce participants more likely to own mobile phones.
  - Non-economic barriers (social norms) can reduce female mobile ownership.

### Uses of ICT and correlates (Table 6 highlights)
- Mobile ownership strongly correlated with online activities:
  - In LIDCs, mobile owners are ~35 percent more likely than national average to pay bills online and ~40 percent more likely to shop, bank online, or receive government transfers via phone (text summary).
- Selected coefficients (LIDC / EME / AE examples):
  - Mobile Owner — Bills:
    - LIDC: 0.017*** [0.002]
    - EME: 0.016*** [0.003]
    - AE: 0.166*** [0.010]
  - Mobile Owner — Shopping:
    - LIDC: 0.013*** [0.002]
    - EME: 0.028*** [0.003]
    - AE: 0.149*** [0.010]
  - Mobile Owner — Banking:
    - LIDC: 0.032*** [0.003]
    - EME: 0.044*** [0.004]
    - AE: 0.167*** [0.010]
  - Mobile Owner — Govt. transfers:
    - LIDC: 0.004*** [0.001]
    - EME: -0.001
    - AE: 0.013*** [0.002]
- Other correlates: education, age, income quintiles, receipt of government transfers, and workforce participation.

### Social dividends from internet adoption (education, time use, labor force participation)
- Empirical strategies: difference-in-differences and IV diff-in-diff (instrument: leave-one-out average internet use in the region).
- Education quality:
  - Increasing internet use from first bin to fifth bin raises average secondary education test scores by between 28 to 44 points.
  - Context: sample average test score 480, minimum 300; using 300 as lowest possible gives effective average = 180; 28-point increase implies 16 percent increase over effective baseline of 180.
- Labor force participation (LFP):
  - Internet use increases LFP by 6 to 10 percentage points over an average of 68 percent when moving from first to fifth bin.
  - Increase larger for women than men.
- Time use:
  - Coefficients suggest unpaid work time decreases for women but remains unchanged for men (evidence not robust).

### Monetized gains from education and labor force participation
- Returns-to-education elasticity used: 0.34 (cross-country log-log relationship between returns to schooling and average test scores).
- Monetized education gains (Schoellman 2012 approach and IPUMS-derived returns):
  - EMEs: average annual gains of 1.1 percent of GDP.
  - LIDCs: average annual gains of 2.3 percent of GDP.
- Labor force participation gains:
  - Average annual wage (Compensation of Employees, PPP constant 2017 USD):
    - EMEs: $16,918.
    - LIDCs: $3,093.
  - Moving bins yields:
    - Average EME: $15 billion in PPP (1.8 percent of GDP PPP).
    - Average LIDC: $0.95 billion in PPP (0.8 percent of GDP PPP).
  - Female LFP accounts for about 58 percent of gains in EMEs and 57 percent in LIDCs.

### Aggregate cost-benefit calculations and present values
- Targets for aggregate calculation (81 EMEs and 51 LIDCs):
  - EMEs: from 72 percent to 90 percent.
  - LIDCs: from 32 percent to 72 percent.
- Summed annual gains from education and increased LFP:
  - LIDCs: approximately $3.7 billion (annual, PPP constant 2017 USD).
  - EMEs: approximately $24 billion (annual, PPP constant 2017 USD).
- Present value of benefits (real discount rate 5 percent; benefits into perpetuity):
  - LIDCs: $3.8 trillion in PPP constant 2017 USD.
  - EMEs: $39 trillion in PPP constant 2017 USD.
- Conservative scenario (returns start after 10 years):
  - LIDCs: $2.3 trillion (PV).
  - EMEs: $23.4 trillion (PV).
- Costs (assumed subsidies perpetual and untargeted):
  - Present discounted costs:
    - EMEs: $4.6 trillion in PPP constant 2017 USD.
    - LIDCs: $4.0 trillion in PPP constant 2017 USD.
  - Targeting reduces costs substantially: PV of targeted subsidies for LIDCs is $2.2 trillion in PPP constant 2017 USD.
- Interpretation:
  - For EMEs: benefits exceed untargeted costs.
  - For LIDCs: case mainly for targeted subsidies (targeted PV cost $2.2 trillion vs untargeted PV cost $4.0 trillion).

### Caveats and downward pressures on subsidy needs
- Subsidies may not be required perpetually as other determinants of internet adoption evolve.
- Network effects from internet adoption can reduce subsidy needs.
- Current calculations use a high price of $16 for LIDCs; expected steep declines in this price could reduce subsidy requirements.

### Policy-relevant takeaways and targeting guidance
- Targeting subsidies is much more cost-effective than untargeted subsidies.
- Suggested target groups (especially in LIDCs):
  - Women (particularly low-education women).
  - Older individuals (≥ 65).
  - Lower education groups.
  - Non-workforce individuals.
  - Lower income quintiles.
- Additional policy considerations:
  - Account for market structure: high concentration may induce supplier price responses to subsidies—proceed conservatively.
  - Complement price subsidies with policies addressing non-price barriers (social norms, digital skills).
  - Monitor trends in mobile internet prices and other determinants to phase out subsidies when no longer necessary.
- Practical targeting feasibility: mobile operators have user information (subscription type, location, spending) that can facilitate eligibility checks.

### Annex I. Potential Determinants of Digital Adoption ....................................................................

### Annex I. Potential Determinants of Digital Adoption

### Location and scope
- Located on page "29" of the source PDF.
- Title: "Annex I. Potential Determinants of Digital Adoption".

### Related annexes and methodological attachments (adjacent in the source)
- Annex II. Data Interpolation, Extrapolation, and Imputation — page "30".
- Annex III. Correlates of Internet Adoption — page "31".
  - Figure AIII.1 Correlates of Internet Use — page "31".
  - Figure AIII.2 Correlates of Internet Access at Home — page "31".
- Annex IV. Details of the Lasso Procedures — page "33".
- Annex V. Validating Parallel Trends Assumption — page "34".
  - Parallel Trends—Simple Difference-in-Differences — page "34".

### Related figures and tables cited in the work (contextually linked to digital adoption)
- Figures (listed in the document):
  - 1. Trends in ICT Expansion by Country Income Groups — page "7".
  - 2. Trends in ICT Expansion by Geographic Regions — page "7".
  - 3. Trends in Internet Coverage and Prices — page "8".
  - 4. Correlates of Internet Adoption — page "10".
  - 5. Strongest Predictors of Digital Adoption — page "12".
  - 6. Trends in Demand Elasticities — page "16".
  - 7. Untargeted and Targeted Subsidies — page "16".
  - 8. Relationship Between Number of Operators and Mobile Data Prices — page "17".
  - 9. Returns to Education Quality — page "26".
- Tables (listed in the document):
  - 1. Correlates of Internet Use — page "13".
  - 2. Demand Estimation — page "15".
  - 3. Estimates of Demand Elasticities — page "15".
  - 4. Summary Descriptive Statistics of Individuals Covered by FINDEX Survey — page "20".
  - 5. Determinants of Mobile Phone Ownership — page "21".
  - 6. Uses of ICT for Various Purposes — page "22".
  - 7. Effects of Internet Use on Education Quality, Labor Force Participation and Unpaid Work — page "25".
  - 8. Monetized Average Annual Social Dividends — page "25".

*Source: wpiea2023065-print-pdf - Annex I. Potential Determinants of Digital Adoption (pages indicated).*

### 9. Total Discounted Costs and Benefits in LIDCs and EMEs ...............................................................

### 9. Total Discounted Costs and Benefits in LIDCs and EMEs

### Major empirical findings on digital adoption
- Digital adoption measure: share of population who report using the Internet in past three months.
- Cross-country disparities (end-2020):
  - AEs: 90 percent Internet use.
  - EMEs: 72 percent Internet use.
  - LIDCs: 32 percent Internet use.
  - Fixed broadband subscriptions: AEs 36 percent, EMEs 14.5 percent, LIDCs 2 percent.
  - Mobile phone ownership: AEs 95 percent, EMEs 80 percent, LIDCs 56 percent.
- Coverage and price patterns:
  - 3G signal coverage and cell phone connectivity: EMEs and LIDCs are close to AEs.
  - 4G connectivity: large gaps, especially between AEs and LIDCs.
  - Effective prices of using the internet (PPP-adjusted): highest in LIDCs despite recent declines.
- Correlates of Internet use (conditional on GDP per capita, within-country variation):
  - Positive correlations: ICT Investment; Coverage of 3G; Mobile phone ownership; Services share of GDP; Secondary education completion; Apps in national languages; Labor force participation; Urbanization; Share of income held by poorest 20 percent; UN E-Gov Score; Gender equality (women empowerment).
  - Negative correlations: Variable costs (price of broadband and mobile data); Poverty and inequality; Rural population; Agriculture and Industry value added.
- Sample and data: 4,124 country-year observations, 169 countries, 1997–2021 (interpolated/extrapolated within a reasonable range).

### Machine learning selection: strongest predictors and policy classification
- LASSO-selected strongest predictors (conditional on GDP per capita and country fixed effects) preserve directions from multivariate regressions.
- Two policy-relevant groups:
  - Fast-moving variables (can be affected short run): broadband and mobile data prices; smart phone ownership (mobile ownership); coverage; use-cases/apps in national languages.
  - Slow-moving variables (move slowly over time): skills (secondary education completion), labor force participation, inequality, urbanization, services share of GDP.
- Policy implication highlighted: fiscal policy can target fast-moving cost variables (e.g., subsidies to lower connection cost) while recognizing slow-moving complements that require structural policies.

### Model of demand for Internet use and price responsiveness
- Consumer utility specification (logit framework) incorporates country fixed effects, year fixed effects, market demand shock, and income-group-specific price sensitivities α_b for b ∈ {L (LIDC), E (EME), A (AE)}.
- Endogeneity handled by control-function approach using leave-one-out regional price averages P−ct as supply shifter; robustness checks use lagged IVs.
- Implied average demand elasticities (reported summary):
  - AE: -0.004
  - EME: -0.26
  - LIDC: -1.22
- Interpretation: demand elasticity is much larger (in absolute value) in LIDCs, moderate in EMEs, and nearly price-inelastic in AEs; demand elasticity decreases with income level and over time.

### Fiscal costs of subsidies to increase Internet use
- Counterfactuals considered: increase Internet use in LIDCs to the level of EMEs and in EMEs to the level of AEs.
- Cost estimates assuming subsidies are untargeted (all users receive discounts on internet fees):
  - EMEs: 0.3 percent of annual GDP.
  - LIDCs: 3.2 percent of annual GDP.
- Cost estimates assuming subsidies are targeted to those deprived of internet:
  - EMEs: 0.1 percent of annual GDP.
  - LIDCs: 1.8 percent of annual GDP.
- Graphical intuition: untargeted subsidies require larger public expense (shaded area under demand curve); perfect targeting reduces public expense but may involve administrative costs and residual deadweight loss.

### Social dividends from faster digital adoption (conservative estimates)
- Labor force participation effects (difference-in-differences and instrumental variables):
  - Increasing Internet use from 10 to 90 percent increases labor force participation by 6 percentage points (average).
  - Women: 7.4 percentage points increase.
  - Men: 4.8 percentage points increase.
- Education quality effects:
  - Same increase in Internet use (10 to 90 percent) increases secondary school test scores by 16 percent on base average.
  - Monetized education gains (using Schoellman 2012 approach and IPUMS International-derived returns):
    - EMEs: average annual gains of 1.1 percent of GDP.
    - LIDCs: average annual gains of 2.3 percent of GDP.
- Gains from increased labor force participation (distributional/aggregate):
  - EMEs: gain of 1.8 percent of GDP.
  - LIDCs: gain of 0.8 percent of GDP.
  - Majority of labor-force-driven gains accrue via increased female labor force participation.
- Time use: weak evidence that Internet use reduces unpaid work time for women.

### Policy considerations and examples
- Rationale for subsidies despite potential distortions:
  - Large social and economic dividends (education, labor participation) can outweigh fiscal costs when properly targeted.
  - Internet adoption exhibits positive network externalities; subsidization may move adoption closer to social optimum.
  - Targeting feasibility: mobile operators have user information (subscription type, location, spending) that can facilitate eligibility checks.
- Country program examples cited:
  - Colombia: ‘Vive Digital’ increased internet adoption among poorer individuals and combined access with training centers.
  - United States: Affordable Connectivity Program (part of Bipartisan Infrastructure Law, 2021) provides monthly discounts to eligible households plus infrastructure investment.
  - Other examples: Costa Rica’s ‘Connected Homes’ program; UK’s ‘Gigabit Voucher Scheme’.
- Trade-offs:
  - Untargeted subsidies are fiscally costly.
  - Targeting reduces fiscal cost but may require administrative capacity and incurs potential targeting costs.
  - Subsidies induce some deadweight loss because some recipients value internet less than the subsidized price.

*Italic: IMF Working Paper content (chapter excerpt).*

### Section IV provides guidance on targeting subsidies.

### wpiea2023065-print-pdf - Section IV provides guidance on targeting subsidies

### Targeting subsidies: definitions and price-discrimination analogy
- Untargeted subsidies are analogous to no price discrimination (one price for all).
- Targeted subsidies (our version) are analogous to third-degree price discrimination (subsidized price for non-adopters).
- Perfectly targeted subsidies (prices equal to the willingness to pay) are analogous to first-degree price discrimination, with government subsidies making up the difference between willingness to pay (the demand curve) and the actual price (eliminating deadweight loss).

### Cost estimates for raising internet use via subsidies
- To increase internet use for EMEs from 72 to 90 percent (the current level in AEs) will require a 96 percent subsidy (18/72/0.26).
- To increase internet use for LIDCs from 32 to 70 percent (the current level in EMEs) will require a 97 percent subsidy (38/32/1.22).
- If untargeted, other things equal, these subsidies amount to:
  - 0.34 percent of annual GDP for EMEs.
  - 3.2 percent of annual GDP for LIDCs.
- If targeted, the corresponding amounts decline to:
  - 0.07 percent of annual GDP for EMEs.
  - 1¾ percent of annual GDP for LIDCs.

- Assumptions used in these calculations:
  - Average monthly price of mobile internet: $4 in EMEs and $16 in LIDCs measured in terms of PPP (ITU).
  - Population numbers (averages over income group from year 2020): 62.5 million in EMEs and 29.7 million in LIDCs.
  - GDP PPP in constant 2017 dollars: $122 billion for an average LIDC and $827 billion for an average EME.
  - To calculate subsidies: product of number of people given subsidies, the level of subsidy in percentage terms, and twelve times the monthly cost of using mobile internet.

### Market structure and supply-side responses
- Regression of price of mobile high basket on number of operators (controlling for country income-group fixed effects and log population) suggests a negative but weak relationship between number of operators and mobile internet price.
- Implication: market concentration likely affects equilibrium outcomes for mobile internet; price subsidies may elicit price responses from suppliers, especially in highly concentrated markets.
- Policy stance in analysis: proceed conservatively and err on the side of caution due to possible supplier responses.

### Composition of the digital divide (who could be targeted)
- Data source: World Bank FINDEX survey, 2017, ~1,000 individuals per country, sampling weights; summary Table 4.
- Cross-country patterns (simple averages):
  - LIDCs have lower mobile phone ownership and fewer debit card owners than EMEs and AEs.
  - LIDCs lag on internet-based activities: internet banking, online bill payments, e-commerce, and ability to get emergency funds.
- Adoption promising aspects in LIDCs:
  - Sending and receiving domestic remittances are highest in LIDCs; almost half of these transactions happen via phones.
  - About 16 percent of government transfers in LIDCs are received via phones—higher than EMEs and AEs.
- Sample summary statistics (Table 4; data from 2017; Standard Deviations in parentheses):
  - Owns Mobile Phone: LIDC 0.660 (0.474); EME 0.847 (0.360); AE 0.935 (0.247).
  - Has Debit Card: LIDC 0.125 (0.331); EME 0.431 (0.495); AE 0.848 (0.359).
  - Used Phone/Internet for Banking: LIDC 0.062 (0.241); EME 0.181 (0.385); AE 0.588 (0.492).
  - Online Bill Payments: LIDC 0.054 (0.225); EME 0.150 (0.357); AE 0.559 (0.497).
  - Bought Online: LIDC 0.036 (0.186); EME 0.144 (0.351); AE 0.544 (0.498).
  - Received Government Transfers: LIDC 0.066 (0.248); EME 0.132 (0.338); AE 0.277 (0.448).
  - Sent Domestic Remittance: LIDC 0.226 (0.418); EME 0.145 (0.352); AE 0.042 (0.201).
  - Sent Remittance by Phone: LIDC 0.112 (0.315); EME 0.033 (0.179); AE 0.005 (0.070).
  - Number of Countries / Observations reported: e.g., LIDC Observations 426,437; EME Observations 212,034; AE Observations 56,740 (as shown in Table 4).

### Within-country digital divide: determinants of mobile ownership
- Regression specification: Owns Mobileict = β0 + β1 Xict + γc + δt + εict (country and year fixed effects; cluster SEs at country level).
- Key regression results (Table 5; data from 2017):
  - Female coefficients on mobile ownership:
    - LIDC: -0.090*** [0.005]
    - EME: -0.062*** [0.004]
    - AE: -0.062*** [0.006]
  - Secondary Education increases mobile ownership:
    - LIDC: 0.168*** [0.007]
    - EME: 0.083*** [0.004]
    - AE: 0.019*** [0.005]
  - Tertiary Education increases mobile ownership:
    - LIDC: 0.214*** [0.016]
    - EME: 0.091*** [0.006]
    - AE: 0.035*** [0.006]
  - Female × Secondary Education and Female × Tertiary Education are positive and significant across income groups.
  - Age effects:
    - Age ≤ 20 and Age between 20 and 65 both positively associated with ownership relative to Age ≥ 65.
  - Income quintiles positively associated with ownership in monotonic fashion (2nd through 5th quintiles).
  - In Workforce: positive association (LIDC 0.069*** [0.005]; EME 0.058*** [0.003]; AE 0.033*** [0.003]).
  - Dependent Variable Means: LIDC 0.69; EME 0.85; AE 0.95.
  - R^2: LIDC 0.181; EME 0.161; AE 0.104.
- Interpretation:
  - Conditional on controls, women are less likely to own mobile phones than men across income groups—gaps are particularly large in LIDCs (e.g., less than primary educated women in LIDCs are 10 percent less likely to own a mobile than comparable men).
  - Mobile ownership increases monotonically with education; older individuals (≥ 65) are less likely to own mobile phones.
  - Richer individuals and workforce participants have higher likelihood of owning a mobile phone.
  - Non-economic barriers (e.g., social norms) can reduce female mobile ownership.

### Uses of ICT and correlates (Table 6)
- Mobile ownership strongly correlated with online activities:
  - In LIDCs, mobile phone owners are 35 percent more likely than national average to pay bills online and approximately 40 percent more likely than national average to shop, bank online, or receive government transfers using a phone (as summarized in text).
- Selected coefficients (Panel structure, LIDC / EME / AE columns):
  - Mobile Owner effects on Bills/Shopping/Banking/Govt. (example entries):
    - LIDC Bills: 0.017*** [0.002]; Shopping: 0.013*** [0.002]; Banking: 0.032*** [0.003]; Govt.: 0.004*** [0.001]
    - EME Bills: 0.016*** [0.003]; Shopping: 0.028*** [0.003]; Banking: 0.044*** [0.004]; Govt.: -0.001
    - AE Bills: 0.166*** [0.010]; Shopping: 0.149*** [0.010]; Banking: 0.167*** [0.010]; Govt.: 0.013*** [0.002]
- Other correlates: education, age, income quintiles, receipt of government transfers, and workforce participation show patterned associations with online uses.

### Social dividends from internet adoption (summary)
- Three key economic outcomes assessed: education quality, time use (percent of 24 hours spent doing unpaid work), and labor force participation (LFP).
- Empirical strategies: difference-in-differences and IV diff-in-diff (instrument: leave-one-out average internet use in the region).
- Main findings (Table 7 and text):
  - Robust positive effect of internet use on secondary school test scores: increasing internet use from first bin to fifth bin increases average secondary education test scores between 28 to 44 points.
    - Context: sample average test score 480, minimum 300; using 300 as lowest possible, effective average = 180; an increase of 28 points implies a 16 percent increase over the effective baseline of 180.
  - Internet use increases LFP by 6 to 10 percentage points over an average of 68 percent when moving from first to fifth bin.
    - Increase in LFP is larger for women than men.
  - Time use: coefficients suggest time spent doing unpaid work decreases for women but remains unchanged for men (not robust).
- Monetized annual gains (back-of-the-envelope, Table 8 and text):
  - Using conservative results from Table 7, Panel B; increasing internet use:
    - For an average EME: annual gains of $9 billion (1.1 percent of GDP).
    - For an average LIDC: annual gains of $2.79 billion (2.3 percent of GDP).
  - Using returns-to-education elasticity of 0.34 (from cross-country log-log relationship between returns to schooling and average test scores).
- Labor force participation gains monetization (steps and results):
  - Average annual wage of a labor force participant (Compensation of Employees data):
    - EMEs: $16,918 in PPP constant 2017 US dollars.
    - LIDCs: $3,093 in PPP constant 2017 US dollars.
  - Moving a LIDC from internet-use bin 2 to bin 5, and an EME from bin 4 to bin 5, yields annual gains:
    - Average EME: $15 billion in PPP (1.8 percent of GDP PPP).
    - Average LIDC: $0.95 billion in PPP (0.8 percent of GDP PPP).
  - Increase in female LFP accounts for about 58 percent of these gains for EMEs and 57 percent for LIDCs.

### Cost-benefit of subsidies and net benefits
- Aggregate estimates for increasing internet usage to target levels (calculations for 81 EMEs and 51 LIDCs):
  - Targets: EMEs from 72 percent to 90 percent; LIDCs from 32 percent to 72 percent.
  - Summed annual gains from education quality improvements and increased LFP:
    - LIDCs: approximately $3.7 billion (annual, PPP constant 2017 US dollars).
    - EMEs: approximately $24 billion (annual, PPP constant 2017 US dollars).
  - Present value of benefits (real discount rate 5 percent; benefits into perpetuity):
    - LIDCs: $3.8 trillion in PPP constant 2017 US dollars.
    - EMEs: $39 trillion in PPP constant 2017 US dollars.
  - Conservative scenario: returns start after 10 years:
    - LIDCs: $2.3 trillion (PV).
    - EMEs: $23.4 trillion (PV).
- Costs (assumed subsidies offered perpetually and untargeted):
  - Present discounted costs:
    - EMEs: $4.6 trillion in PPP constant 2017 US dollars.
    - LIDCs: $4.0 trillion in PPP constant 2017 US dollars.
  - Targeting reduces costs substantially: present value of targeted subsidies for LIDCs is $2.2 trillion in PPP constant 2017 US dollars.
- Interpretation:
  - Case for internet subsidies appears clear for EMEs (benefits exceed untargeted costs).
  - For LIDCs, case exists mainly for targeted subsidies (targeted PV cost $2.2 trillion vs untargeted PV cost $4.0 trillion).
- Caveats and downward pressures on subsidy needs:
  - Subsidies may not be required perpetually as other determinants of internet adoption evolve.
  - Network effects from internet adoption can reduce subsidy needs.
  - Current calculations use a high price of $16 for LIDCs; this price is expected to decline steeply according to trends.

### Policy-relevant takeaways and targeting guidance
- Targeting subsidies is much more cost-effective than untargeted subsidies; targeted subsidies reduce required public expenditure substantially (examples above).
- Target groups suggested by the digital-divide analysis:
  - Women (particularly low-education women), older individuals (≥ 65), lower education groups, non-workforce individuals, and lower income quintiles—especially in LIDCs.
- Additional policy considerations:
  - Account for market structure: high concentration may induce supplier price responses to subsidies—proceed conservatively.
  - Complement price subsidies with policies addressing non-price barriers (e.g., social norms restricting female ownership, digital skills for older populations).
  - Monitor trends in mobile internet prices and other determinants of adoption to phase out subsidies when no longer necessary.

*Source: wpiea2023065-print-pdf - Section IV provides guidance on targeting subsidies. Canonical URL: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023065-print-pdf.pdf*

### Annex I. Potential Determinants of Digital

### Annex I. Potential Determinants of Digital Adoption

### Variables and Measurement
- Name of the Variable — Source — Unit
  - ICT Investment — World Bank — % GDP
  - Population Density — World Bank — 100 per sq km
  - Electricity Access — World Bank — %
  - Log Fixed Broadband — ITU — $ PPP
  - Log Mobile High Package — ITU — $ PPP
  - Log Mobile Low Package — ITU — $ PPP
  - Industry Value Added — World Bank — % GDP
  - Services Value Added — World Bank — % GDP
  - Agriculture Value Added — World Bank — % GDP
  - GDP per Capita — World Bank — $1,000
  - Education Completion (Secondary) — World Bank — %
  - Education Completion (Primary) — World Bank — %
  - Women Empowerment — World Bank — %
  - Inequality — World Bank — GINI
  - Poverty at $1.9 — World Bank — Head-Count Ratio ($1.9)
  - Rural Population — World Bank — %
  - Lowest 20% Income — World Bank — %
  - 2nd Lowest 20% Income — World Bank — %
  - 3rd Lowest 20% Income — World Bank — %
  - 2nd Highest 20% Income — World Bank — %
  - Highest 20% Income — World Bank — %
  - Labor Force Participation Rate — World Bank — %
  - Population Covered 3G (%) — ITU — %
  - Handset Price — GSMA — % Monthly GDP
  - Gender Parity — GSMA — %
  - UN E-Government Score — UN — Score (1-100)
  - Apps in National Language(s) — GSMA — %
  - Mobile Ownership — GSMA — % Population

### Data Interpolation, Extrapolation, and Imputation
- Study period covered: 1997-2021.
- Method for filling missing World Bank World Development Indicators:
  - Linearly interpolate and extrapolate using the gradient from the two most recent years available for each variable.
  - Example procedure: If X has values in 2011, 2014 and 2017, use 2014 and 2017 to linearly interpolate 2015 and 2016 and to extrapolate for 2018 and later; use 2011 and 2014 to interpolate 2012 and 2013 and to extrapolate for years before 2011.
- Handling of implausible extrapolations:
  - All negative extrapolated values are replaced with missing values.
  - For variables measured in percentages, all extrapolated values above 90 percent are replaced with missing values.
- Approach to enable multivariate regressions despite remaining missing values:
  - Replace missing values with average values (means) of each variable.
  - Generate a corresponding dummy variable that takes value one for observations imputed using means.
  - Regression specifications include these dummy variables as controls to partial out imputed values.

### Correlates of Internet Adoption (Notes from Figures)
- Figure AIII.1 (Correlates of Internet Use):
  - Dependent measure: percent of population.
  - Reported quantities: regression coefficients β1 from equation 1.
  - Estimation details: Income Group x Year fixed effects and standard errors clustered at country level.
  - Controls include GDP per capita.
- Figure AIII.2 (Correlates of Internet Access at Home):
  - Dependent measure: percent of population.
  - Reported quantities: regression coefficients β1 from equation 1.
  - Estimation details (figure notes vary): Country fixed effects and standard errors clustered at country level; other notes indicate Income Group x Year fixed effects and standard errors clustered at country level.
  - Controls include GDP per capita.

### Details of the Lasso Procedures
- Software: rlasso in STATA package ‘LASSOPACK’.
- LASSO estimator β̂ solves:
  - minβ 1/N RSS + (λ/N) ‖ψ × β‖1
  - Where RSS = ∑(yi − xi′β)2; β is p-dimensional parameter vector; λ is the overall penalty level; ‖.‖1 denotes the L1-norm (∑|ai|i); ψ is a p by p diagonal matrix of predictor-specific penalty loadings (rLASSO treats ψ as a row vector); N is the number of observations.
- Partialed out variables and controls in LASSO estimation:
  - GDP per capita
  - Country fixed effects
  - All dummy variables for imputed values corresponding to included predicting variables
- Kernel and bandwidth:
  - Default Bartlett kernel with bandwidth 11 (order) has been used.

### Validating Parallel Trends Assumption
- Method referenced: Parallel Trends—Simple Difference-in-Differences.
- Figures and results labeled "Source: Authors’ calculations" accompany the parallel trends validation.

*IMF Working Paper — Annex I, II, III, IV, and V excerpts from Determinants and Social Dividends of Digital Adoption (Working Paper No. WP/2023/065).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023065-print-pdf.pdf_