## 2015.   Notably,  the  introduction  of  3G  mobile  internet  at  the  country  level  occurred  in  2011,

## Source details

**Canonical URL:** [2015.   Notably,  the  introduction  of  3G  mobile  internet  at  the  country  level  occurred  in  2011,](https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024070-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2024/english/wpiea2024070-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2024/english/wpiea2024070-print-pdf.pdf.json)

---

### Executive summary and research design
- Unit of observation: municipality (sector) level in Rwanda (5 provinces, 30 districts, 416 sectors).
- Time variation: 3G mobile internet introduced at the country level in 2011; 3G accounts for most variation over 2008–2015.
- Cross-sectional identification: instrumental variables — lightning strike frequency (Global Hydrology Resource Center) and incidental coverage (Björkegren (2019)).
- Mechanism data: Land Tenure Regularization (LTR) program administrative data on land certificates, transactions, and mortgaged lands (collaboration with Rwanda Land Management and Use Authority).
- Main datasets:
  - Credit Reference Bureau (CRB) loan-level data: monthly reporting January 2008–December 2015, aggregated yearly.
  - GSMA/C​ollins Bartholomew 3G coverage maps (spatial resolution 1 km2 to 15–23 km2).
  - GHRC lightning data (grid resolution 0.1) and Björkegren incidental coverage maps.

### Main findings: mobile internet and the credit market
- Reallocation of borrowing:
  - 3.4% increase in the probability of having a loan from a commercial bank (response to a one standard deviation increase in 3G coverage).
  - 3.6% decrease in the probability of having an MFI loan (response to a one standard deviation increase in 3G coverage).
  - 1.1% higher probability that an MFI customer switches to a commercial bank (one standard deviation increase in 3G coverage).
- Intensive-margin borrower characteristics:
  - A one standard deviation increase in 3G coverage associated with a 6.5% increase in outstanding loan amount among bank borrowers.
  - No aggregate adverse effect on interest rates for bank borrowers.
  - For switchers (MFI → bank), a significant reduction in lending rates observed, especially for switchers in the upper tail of the interest rate distribution.

### Mechanism: land titles, collateral, and mortgages
- LTR program digital registry: database had information on 10.4 million land parcels by 2012.
- Mobilization and digital outreach: National Land Authority Twitter account introduced in March 2011; approximately 110,000 people mobilized during data collection and corrections.
- Evidence on collateral channel (responses to a one standard deviation increase in 3G coverage):
  - 16 percentage points higher likelihood that the municipality experiences a disbursement of land titles exceeding 50% of the total in the first year of the reform.
  - 14% increase in the total number of transactions.
  - 1.7 percentage point higher share of mortgaged lands.
  - 0.6 percentage point increase in the probability of receiving a collateralized bank loan.
  - 0.8 percentage point increase in the probability of receiving a mortgage from a commercial bank.
- Mediation (Sobel-Goodman and related tests):
  - Approximately 35% of the effect of mobile internet on the probability of receiving a bank loan is indirect and mediated by land certificates; 65% is direct.
  - When the dependent variable is probability of receiving a collateralized loan: direct effect falls to 20% (borderline significant) and mediated effect by land titles is 80% of total impact.
  - Sequential g-estimation (Acharya et al., 2016) corroborates results and indicates two-thirds of total effect on bank loan probability is not mediated through land titles in that specification.

### Institutional and contextual background (Rwanda)
- Financial inclusion and credit markets:
  - Financial inclusion increased from 48% to 93% between 2008 and 2020.
  - Share of adult population using commercial banks increased from 14 to 36% between 2008 and 2020 (Access to Finance Rwanda, 2020).
  - MFIs historically key in rural/peri-urban credit using group lending.
- Mobile internet and ICT:
  - Rwanda ranked third in Africa in percentage population covered by 3G in 2015 (GSMA, 2015).
  - RURA reported 3.8 million mobile internet subscriptions among a population of almost 11.5 million.
  - Rwanda had the cheapest price per gigabyte in East Africa as a percentage of per capita income.
- LTR program:
  - Launched in 2009; by 2013 all financial institutions were connected to the digital registry.
  - Mobile internet used directly and indirectly to promote the program and provide updates.

### Data, identification, and estimation
- CRB borrower sample (final cleaned sample for local currency lending activities of banks and MFIs):
  - 113,897 unique individuals; 337 municipalities (out of 416).
- Individual loan-level variables: loan amount and price, borrower municipality, borrower age/gender/marital status.
- Instruments and 2SLS setup:
  - Instruments: Z_light_m × post-2010 and Z_incid_m × post-2010; first stage and second stage LPM with borrower and year fixed effects.
  - First-stage evidence: Std Lightning coefficients negative and significant; Std Incidental coefficients positive and significant. F-statistics reported in Table 2 include values 10.64, 23.17, 17.74, 18.49, 11.81, 16.22 across columns.
- Baseline regression:
  - Y_imt = v + β1 3G_Coverage_mt + α_i + φ_t + ε_imt, with 3G_Coverage_mt standardized and LPM estimated with robust standard errors clustered at municipality level.

### Robustness, heterogeneity, and alternative specifications
- Heterogeneity:
  - Effects larger in urban and wealthier areas and where a physical bank branch is present, though some differences not statistically significant.
  - No discernible heterogeneity by borrower gender, marital status, or age.
- Robustness checks (selected exact findings):
  - Aggregation to bank-municipality-year level: one standard deviation increase in 3G coverage associated with a 13% increase in number of loans, a similar increase in number of switchers, 33% increase in total loan outstanding, and no significant effect on average interest rates (Table B1).
  - All-borrowers sample (Table B2): Probability of Bank Loan 0.033*** (0.012); MFI Loan -0.035*** (0.013).
  - Dummy 3G (Table B4): Probability of Bank Loan 0.094*** (0.036); MFI Loan -0.110*** (0.035).
  - Weighted 3G coverage by population density / total population yields larger magnitudes (Table B5).
  - Replacing instruments with predetermined geographic features (altitude; percentage covered by wood, water, sand) confirms main results with SW F-statistic above 100 in Table B8 (example SW F-statistic: 861.54).
  - Reduced-form regressions for instruments and geographical covariates align with main specifications (Table B9).
  - Inclusion of municipality time trends and controls for U-SACCOs and fiber rollout does not overturn main coefficients (Tables B10–B12).
  - Fixed broadband falsification: fiber rollout (completed end-2010) controlled for; results on collateral channel robust (Table B13).

### Event timing and dynamics
- Event-study: effects on bank loans are lagged, becoming significant after two years from 3G introduction; absence of pre-trends supports identification (Figure A1 description).

### Key numerical summaries and diagnostics (selected exact figures)
- Sample moments (Table 1, Panel A):
  - Probability of Any Loan: mean 0.280, sd 0.45.
  - Probability of Bank Loan: mean 0.140, sd 0.34.
  - Probability of MFI Loan: mean 0.160, sd 0.37.
  - Switching borrowers: mean 0.090, sd 0.29.
- Bank loan characteristics (Table 1, Panel B):
  - Outstanding loan (n = 124,806): mean 2011.88, sd 3665.80, 50 693.11, min 16.67, max 27859.83 (units: thousands of Rwandan francs).
  - Interest rate (n = 124,388): mean 19.20, sd 5.73, 50 19.00, min 6.14, max 70.10 (percentage terms).
  - Collateral dummy (n = 128,639): mean 0.12, sd 0.32.
- Main IV 2SLS estimates (Table 3, Std 3G coverage):
  - Bank Loan (IV): 0.034*** (0.012).
  - MFI Loan (IV): -0.036*** (0.012).
  - SW F-statistic: 16.22.
  - Observations: 909,165.
- Switching estimates (Table 4, IV):
  - Probability of Switching from MFI to Bank: 0.011*** (0.003).
  - Probability of Switching from Bank to MFI: -0.011** (0.005).
  - Observations: switching from MFI to Bank 514,372; Bank to MFI 380,974.
- Land transactions and mortgaged land (Table 7 and Table 8, IV and OLS highlights):
  - 15 percentage point higher probability of disbursing at least 50% of titles in first year (Table 7, IV: 0.147* (0.080)).
  - 13% increase in total number of transactions (Table 7, IV: 0.128*** (0.036)).
  - Share of land mortgaged (Table 8, examples): OLS 0.024*** (0.003); IV 0.009*** (0.001).
  - Probability(Bank Collateral) IV: 0.006** (0.003); Probability(Bank Mortgage) IV: 0.008*** (0.002).
- Mediation test (Table 9 exact values):
  - Indirect Effect (Collateral) on Probability(Bank Loan): 0.013**.
  - Direct Effect (Mobile Internet) on Probability(Bank Loan): 0.021**.
  - Total Effect: 0.034**.
  - Proportion Indirect: 37% (Bank Loan); 80% (Bank Collateral).

### Policy implications and contributions
- Technology-law complementarity:
  - Mobile internet amplifies legal reforms (digital land registry) by lowering costs of acquiring and enforcing property rights, thereby facilitating access to collateralized formal credit.
- Policy recommendations:
  - Promote mobile infrastructure and internet connectivity to enable use of formal banking services and collateralization of assets.
  - Tailor interventions to local conditions: effects are larger in urban, wealthier areas and where bank branches exist; extremely rural and poorest areas may need complementary measures.
  - Ensure financial institutions and policymakers support switchers to realize better loan conditions and broaden access to collateral channels.
- Contribution to literature:
  - Documents reallocation from MFIs to banks without net inclusion gains, complements literature on property rights and credit access, and links internet technologies to mortgage lending and household finance in emerging markets.

*Italicized source: wpiea2024070-print-pdf - 2015. Notably, the introduction of 3G mobile internet at the country level occurred in 2011,*

### 2015.   Notably,  the  introduction  of  3G  mobile  internet  at  the  country  level  occurred  in  2011,

### wpiea2024070-print-pdf - 2015. Notably, the introduction of 3G mobile internet at the country level occurred in 2011,

### Executive summary and research design
- The unit of observation is at the municipality (sector) level in Rwanda. A municipality refers to the third level administrative subdivision; there are 5 provinces, 30 districts, and overall 416 sectors.
- The introduction of 3G mobile internet at the country level occurred in 2011, providing time-series variation in internet connectivity. Cross-sectional variation is exploited using instrumental variables: lightning strike frequency (Global Hydrology Resource Center) and incidental coverage (Björkegren (2019)).
- The analysis incorporates the Land Tenure Regularization (LTR) program and novel municipal-level data on the distribution of land certificates, number of transactions, and share of mortgaged lands. Collaboration was with the Rwanda Land Management and Use Authority.

### Main findings: mobile internet and the credit market
- Mobile internet availability prompted a reallocation of borrowing away from microfinance institutions (MFIs) toward commercial banks, without changing overall financial inclusion (likelihood of receiving a loan from any financial institution).
- Exact estimated effects (responses to a one standard deviation increase in 3G coverage):
  - 3.4% increase in the probability of having a loan from a commercial bank.
  - 3.6% decrease in the probability of having an MFI loan.
  - 1.1% higher probability that an MFI customer switches to a commercial bank.
- Borrower characteristics and loan terms among bank borrowers:
  - Borrowers with mobile internet access tend to have larger outstanding loan amounts, with no adverse effects on interest rates.
  - For switchers from MFIs to commercial banks, especially those in the upper tail of the interest rate distribution, a significant reduction in the lending rate is observed.

### Mechanism: land titles, collateral, and mortgages
- The Land Tenure Regularization (LTR) program established a digital registry and created digital records for over 10 million land parcels; by 2012 the database had information on 10.4 million land parcels.
- The LTR program was supported by digital advertising and use of mobile internet (National Land Authority Twitter account introduced in March 2011); a large mobilization involved approximately 110,000 people during data collection and corrections.
- Evidence that mobile internet facilitates acquisition and use of land titles as collateral:
  - A one standard deviation increase in 3G coverage leads to a 16 percentage points higher likelihood that the municipality experiences a disbursement of land titles exceeding 50% of the total in the first year of the reform.
  - A one standard deviation increase in 3G coverage leads to an increase by 14% of the total number of transactions.
  - A one standard deviation higher 3G coverage leads to a 1.7 percentage point higher share of mortgaged lands.
  - The probability of receiving a collateralized bank loan or a loan classified as mortgage increases by 0.6 and 0.8 percentage point, respectively, in response to a one standard deviation increase in 3G coverage.
- Mediation analysis (Sobel-Goodman test) quantifying collateral’s role:
  - Approximately 35% of the effect of mobile internet on the probability of receiving a bank loan is indirect and facilitated by the proxy of land certificates’ availability; 65% is directly attributed to mobile internet.
  - When using the probability of receiving a collateralized loan as dependent variable: the direct effect of mobile internet decreases to 20% (borderline significant) and the effect mediated by land titles accounts for 80% of the total impact.
  - These quantifications indicate a pivotal role of collateral in securing mortgages and steering individuals from MFIs toward commercial banks.

### Institutional and contextual background (Rwanda)
- Credit market and financial inclusion:
  - Historically limited access to financing; MFIs played a key role in rural and peri-urban credit provision, employing group lending methodologies.
  - Financial inclusion increased from 48% to 93% between 2008 and 2020.
  - The share of adult population using services and products from commercial banks increased from 14 to 36% between 2008 and 2020 (Access to Finance Rwanda, 2020).
- Mobile internet and ICT rollout:
  - Rwanda’s Vision 2020 and subsequent policies promoted ICT as a development strategy.
  - In 2015 Rwanda ranked third in Africa in the percentage of the population covered by 3G (after South Africa and Lesotho) (GSMA, 2015).
  - RURA reported 3.8 million mobile internet subscriptions among a population of almost 11.5 million.
  - Affordability: Rwanda had the cheapest price per gigabyte in East Africa when considered as a percentage of per capita income.
- Land Tenure Regularization (LTR) program:
  - Launched in 2009 to regulate ownership and control of lands after post-conflict returnee migration and land tensions.
  - The LTR introduced a land tenure system based on a digital registry; by 2013 all financial institutions in Rwanda were connected to the digital registry.
  - Mobile internet was used directly (information search) and indirectly (channels to other information sources) to promote the program and provide updates.

### Robustness, heterogeneity, and identification
- Heterogeneity:
  - Socioeconomic differences between municipalities drive heterogeneous effects; individual-specific observable traits (gender, marital status, age) show no discernible role, suggesting inclusive technology effects in this context.
- Robustness checks and alternative specifications:
  - Results persist when aggregating to the bank-municipality-year level.
  - Alternative definitions of dependent variables, predictors, and instrumental variables yield similar results.
  - Replacing main instruments (lightning strikes and incidental coverage) with pre-determined physical characteristics (altitude and coverage of wood, water and sand) delivers similar estimates, with higher precision and a first stage F-statistic exceeding 100.
  - Reduced-form estimates for both IV strategies align with main specifications.
  - Results robust to inclusion of municipality time trends and controls for other reforms (Umurenge Savings and Credit Cooperatives (U-SACCOs) and fiber rollout).
  - Mediation results robust to a falsification test focusing on fiber-optic availability and to a sequential g-estimation procedure that produces comparable quantifications (Acharya et al., 2016).

### Implications and contribution to literature
- Technology-law complementarity:
  - The paper shows how mobile internet interacts with legal reforms (digital land registry) to reduce costs of acquiring and enforcing property rights, facilitating access to formal credit.
- Relation to prior findings:
  - Echoes Breza and Kinnan (2021) on shifts from MFIs to banks without net inclusion gains.
  - Connects with literature on property rights, collateral, and credit access (La Porta et al., 1998; Djankov et al., 2007; Besley et al., 2012; Djankov et al., 2020).
  - Contributes to literature on technology’s role in mortgage lending and household finance in emerging markets (Fuster et al., 2019; Badarinza et al., 2019) and to studies on internet technologies’ societal impacts (Guriev et al., 2021; Manacorda and Tesei, 2020).

*Italicized source: wpiea2024070-print-pdf - 2015. Notably, the introduction of 3G mobile internet at the country level occurred in 2011, (canonical PDF).*

### 3.1    Data

### 3.1    Data

### Data sources and sample
- Comprehensive loan-level data from the Credit Reference Bureau (CRB), regulated by the National Bank of Rwanda (NBR).
- Coverage: loans provided by all supervised credit institutions in Rwanda, including commercial banks, microfinance institutions, and savings and credit co-operatives (SACCOs).
- Reporting frequency: monthly, from January 2008 to December 2015.
- No minimum loan size requirement.
- Aggregation used in analysis: yearly level.
- Final cleaned borrower sample (local currency lending activities of banks and MFIs):
  - 113,897 unique individuals
  - 337 municipalities
- Borrowers are tracked over time and across lenders via a unique numerical code.
- Note: total municipalities in Rwanda = 416; missing municipalities in the sample result from inability to differentiate between municipalities with the same name but in different districts in the credit bureau data. Observations with missing borrower age are excluded to avoid including borrowers who may be under the age of 16 in the balanced panel analysis.

### Individual loan-level variables (baseline analysis)
- Focus: loans granted to individuals.
- Available per-loan information includes:
  - loan amount and price
  - borrower’s location (down to municipality level)
  - borrower characteristics: age, gender, marital status

### Mobile internet coverage data
- Source: GSM Association (GSMA) via a partnership with Collins Bartholomew.
- Time coverage: yearly geolocated information from 2008 to 2015.
- Technology distinctions: availability of 2G, 3G, and 4G; 3G coverage accounts for most variation over the sample period.
- Spatial resolution for 3G penetration measurement:
  - between 1km2 on the ground (for high-quality GIS vector submissions)
  - to 15-23km2 for submissions based on antenna locations and radius of coverage (GSMA, 2015).
- Mobile adoption mapped at municipality-year level.
- GSM network market share context: GSM network is the dominant standard in Africa and covers around 96% of the market share. These data come from submissions made directly by mobile operators for constructing roaming coverage maps for end users.

### Instrumental variables for identification
- Two IVs constructed using external datasets to address endogeneity and establish causality between 3G technology and access to credit.

1) Lightning frequency (Z_light)
- Source: Global Hydrology Resource Center (GHRC).
- Data collection platforms: spaceborne Optical Transient Detector (OTD) on Orbview-1 and the Lightning Imaging Sensor (LIS) onboard TRMM.
- Detection efficiencies:
  - LIS: 88% during the night, 69% at noon.
  - OTD: 52% during the night, 37% at noon.
- Data released as part of the LIS/OTD Gridded Lightning Climatology Data Collection.
- Processing: time and lightning are summed per grid cell and scaled for sensor precision, providing daily, monthly, seasonal, or yearly data.
- Resolution used: grid resolution of 0.1 (improvement from 0.5 in prior work).
- Instrument definition in empirical strategy: interaction between time-invariant average of lightning strikes Z_light_m and a dummy post-2010 (post-2010 = 1 after 2010).

2) Incidental coverage (Z_incid)
- Concept: fictitious mobile coverage that would have resulted if mobile operators had built towers along the entire network of power lines; captures variation from interaction of geography and preexisting electric grid.
- Construction follows Björkegren (2019) methodology.
- Motivation: operating towers connected to the electric grid is more cost-effective; topography (visibility from 35 meters above the tower) affects realized coverage.
- Instrument definition in empirical strategy: interaction between time-invariant average incidental coverage Z_incid_m and a dummy post-2010.
- Incidental coverage data kindly provided by Daniel Bjorkegren.

- Both IVs are time-invariant measures interacted with a post-2010 dummy to exploit differential 3G adoption after the 3G introduction (3G technology officially introduced in 2010; during the first year, less than 8% of the country was connected).
- Empirical evidence (described in the source):
  - Municipalities below and above median lightning strikes had zero 3G coverage before 2011, after which trends diverge monotonically.
  - Municipalities below and above median incidental coverage had zero 3G coverage before 2011, after which trends diverge monotonically.
- Balance checks (Table A1 in appendix) compare pre-2011 means across quartiles of lightning strikes and incidental coverage; results are described as largely reassuring for instrument validity with some small geographical correlations controlled for in robustness checks.

### Additional datasets for mechanisms and controls
- Rwanda Land Management and Use Authority land dashboard:
  - Information on land certificates, transactions, and mortgaged lands.
  - Proprietary municipality-level land title data obtained to analyze relationship between land collateral and bank credit access (district-level titles are publicly available; municipality-level proprietary data used for analysis).
- Municipality-level controls:
  - National Institute of Statistics (NISR)
  - Integrated Household Living Conditions Survey (EICV)
  - Focus: indicators related to local economic and financial activities, particularly prior to 3G introduction (before 2011).

### Summary statistics and variable panels (described)
- Table 1 (described):
  - Panel A: access to credit — probability of having an outstanding loan; separate probabilities for commercial bank and MFI loans; switching borrowers.
  - Panel B: yearly average bank credit relationship characteristics — outstanding amount, principal amount, interest rate, probability that the loan is secured.
  - Panels C and D: summary statistics for main predictor (3G mobile coverage) and instruments (lightning frequency, incidental coverage).

### Key methodological notes
- Analysis unit: balanced panel at borrower-year level for access outcomes (the panel is “almost” balanced because observations for borrowers younger than 16 are removed).
- When analyzing bank loan characteristics (intensive margin), an unbalanced borrower-bank-year panel is used to allow inclusion of bank-time fixed effects.
- Baseline regression (linear probability model, LPM):
  - Y_imt = v + β1 3G_Coverage_mt + α_i + φ_t + ε_imt
  - Y_imt is either:
    - binary: indicator of outstanding loan with any institution, or with bank, or with MFI; or switching to bank/MFI by time t or before;
    - continuous: amount of outstanding loan or interest rate for loan characteristics specifications.
  - 3G_Coverage_mt is the standardized measure of percentage mobile internet coverage in municipality m at time t.
  - Controls: borrower fixed effects (α_i) and time fixed effects (φ_t).
  - Estimation: LPM with robust standard errors clustered at municipality level.
  - Rationale for LPM: consistent interpretation with Local Average Treatment Effect (LATE) and avoids incidental parameters problem for maximum likelihood estimators.

- 2SLS setup:
  - First stage: 3G_Coverage_mt = q + δ1 Z_light_mt + δ2 Z_incid_mt + θ_i + ξ_t + ε_imt
  - Second stage: Y_imt = v + β1 \3G_Coverage_mt + α_i + φ_t + ε_imt
  - Instruments: Z_light_mt = Z_light_m × post-2010; Z_incid_mt = Z_incid_m × post-2010.
  - Identification assumption: correlations between instruments and municipality characteristics remained unchanged at the time of 3G introduction; effect identified from change in impact of lightning/incidental coverage post-2010.

- First-stage evidence (Table 2 described):
  - Columns 1–3: municipality-year sample (one observation per municipality-year).
  - Column 1: lightning strikes as sole instrument — coefficients negative and statistically significant.
  - Column 2: incidental coverage as sole instrument — coefficients positive and statistically significant.
  - Column 3: both instruments included — both significant and F-statistics generally high, surpassing typical threshold values indicating instrument relevance.
  - Columns 4–6: re-estimation on sample of all borrowers in credit registry — results described as consistent.

- Robustness:
  - Alternative IV specification considered in robustness: interacting Z_light_m and Z_incid_m with a linear time trend t (described as adopted in robustness section).

*Source: IMF working paper section 3.1 Data (wpiea2024070-print-pdf).*

### 3.6 percentage point decrease in the probability of having a loan from a MFI (column 6).

### wpiea2024070-print-pdf - 3.6 percentage point decrease in the probability of having a loan from a MFI (column 6).

### Heterogeneous impact of mobile internet on access to credit
- 3.6 percentage point decrease in the probability of having a loan from a MFI (column 6).
- The impact of mobile internet on access to credit is heterogeneous: it positively affects access to formal banking, while negatively affecting access to microfinance.
- The results imply mobile internet contributes to financial graduation and the move towards commercial banks:
  - On the extensive margin, potential borrowers endowed with the new internet technology prefer commercial banks to MFIs.
  - On the intensive margin, incumbent borrowers transition to formal banking.
- Hansen J-statistic p-values for specifications in columns 5 and 6: 0.39 and 0.94, respectively (overidentifying restrictions on the instruments are not rejected).

### Direct evidence on transition from MFIs to banks (Section 4.2)
- Four borrower types identified: only banks; only MFIs; initial MFI then switch to bank; initial bank then access MFI.
- Sample split into “First MFI” and “First bank”; OLS and 2SLS regressions estimated with Probability(of Switching)imt as dependent variable.
- Key estimates (Table 4):
  - A one standard deviation increase in 3G mobile coverage is associated with a 1.1 percentage point increase in the probability of switching to banks (column 2, borrowers with initial MFI).
  - A one standard deviation increase in 3G mobile coverage is associated with a 1.1 percentage point decrease in the probability of switching to MFIs (column 4, borrowers with initial bank).
- Conclusion: mobile internet has a positive effect on the intensive margin of transition from microfinance to formal banking; incumbent MFI borrowers increase probability of moving to commercial banks.

### Timing and heterogeneity in access to bank credit (Section 4.3)
- Event-study specification centered on relative year from the introduction of 3G; year before arrival of 3G is the omitted category.
- Event-study findings:
  - The effect of mobile internet on bank loans is lagged and takes time to materialize.
  - The effect becomes significant after two years following the introduction of 3G and is notable in magnitude.
  - Absence of pre-trends supports identification strategy.
- Socio-economic heterogeneity (Table 5, panel A):
  - Effect of 3G coverage appears larger in urban and wealthier areas, and where a physical bank branch is present.
  - Differences are not statistically significant; effect may be negligible in extremely rural and the poorest areas.
- Borrower characteristic heterogeneity (Table 5, panel B):
  - Effect of 3G coverage on probability of obtaining a commercial bank loan is homogeneous across gender, marital status (single versus married), and age (young versus adult).
  - Differences in magnitudes are not statistically significant.

### Bank loan characteristics and intensive margin effects (Section 4.4)
- Analysis at the individual borrower-bank-year level for borrowers who accessed bank loans.
- Main findings (Table 6, columns 1 and 2):
  - A one standard deviation increase in 3G coverage is associated with an increase in outstanding loan amount by 6.5%.
  - No evidence of an effect on interest rates at the aggregate borrower level.
- Heterogeneous effects for switchers (borrowers who initially had a MFI loan and switched to a commercial bank), using interaction with Switcheri (Equation (5)):
  - Switchers experience a relative increase in loan outstanding and a significant decrease in the average interest rate (Table 6, columns 3 and 4).
  - Borrowers keeping relationships with MFIs show no effect from 3G on loan conditions.
- Implication: mobile internet enhances borrowing opportunities and loan terms for switchers without adverse effects on loan prices for bank borrowers overall.

### Complementarity between technology and law — the collateral channel (Section 4.5 and 4.5.1)
- Hypothesis: technology reduces costs of accessing and utilizing property rights, enabling land titles to be leveraged as collateral and facilitating access to banks.
- Rwanda Land Tenure Regularization (LTR) program:
  - By 2012, the main database had accumulated information on more than 10.4 million land parcels.
  - Registration speed improved dramatically: registration time fell from over 300 days to less than 40 days in under five years.
  - Mobile internet played a crucial role in informing people and disseminating the reform’s procedures and technical aspects.
- Five pieces of evidence used to support the collateral channel (summarized):
  1. Speed of registration of land transactions and relevance of LTR in reducing transaction costs.
  2. Positive relationship between 3G coverage and number of land title transactions.
  3. Association between 3G coverage and the amount of mortgaged land parcels.
  4. Relationship between 3G coverage and bank credit collateral.
  5. Quantification of mediation of mobile internet’s effect on banking outcomes through the collateral channel (Sobel-Goodman mediation test referenced).
- Empirical results (Tables 7 and 8):
  - Table 7 (municipality-level, period ending 2013; 2SLS results in columns 3 and 4):
    - A one standard deviation increase in 3G coverage is associated with a 15 percentage point higher probability of disbursing at least 50% of the titles in the first year of the reform.
    - A one standard deviation increase in 3G coverage is associated with a 13% increase in the total number of transactions.
  - Table 8 (mortgaged land and collateral outcomes):
    - A one standard deviation increase in 3G coverage is associated with a 1.7 percentage point increase in the share of land parcels used as collateral (columns 1–2).
    - A one standard deviation increase in 3G coverage is associated with a 0.6 percentage point increase in the probability of a bank loan being backed by collateral (column 5).
    - A one standard deviation increase in 3G coverage is associated with a 0.8 percentage point increase in the probability of the loan being a mortgage issued by a commercial bank (column 6).
- The evidence supports mobile internet promoting land reform implementation and increasing access to collateralized bank credit.

### Policy implications and recommendations
- Promote development of mobile infrastructure and improving internet connectivity to create an enabling environment for individuals to access and utilize formal banking services.
- Consider local economic conditions:
  - Effects of mobile internet on accessing commercial bank loans appear larger in urban, wealthier areas and where physical bank branches exist; interventions may need to be tailored for extremely rural and the poorest areas.
- Recognize the complementarity between technology and legal/property reforms:
  - Mobile internet amplifies the impact of land titling reforms (LTR) by facilitating dissemination and uptake, which in turn increases access to collateralized loans and mortgages.
- Financial institutions and policymakers should consider targeted measures to ensure switchers benefit from better loan conditions and that collateral channels are accessible to broaden formal credit access.

*Italic: Source — wpiea2024070-print-pdf - 3.6 percentage point decrease in the probability of having a loan from a MFI (column 6).*

### 9.  Consistently with our story,  we find that approximately 35% of the  probability of receiving

### 9.  Consistently with our story,  we find that approximately 35% of the  probability of receiving

### Mediation and channels linking mobile internet, collateral, and bank credit
- Approximately 35% of the probability of receiving a bank loan is due to an indirect effect facilitated by the acquisition of the collateral, whereas 65% is directly attributed to mobile internet.
- The “direct effect” of mobile Internet is a residual effect and captures many potential channels, including:
  - better job opportunities (Hjort and Poulsen, 2019),
  - higher income (Bahia et al., 2023; Calderone et al., 2018),
  - lower information frictions (Gupta et al., 2023).
- When the dependent variable is the probability of receiving a collateralized loan:
  - the direct effect of mobile internet decreases to 20% and is borderline significant,
  - the indirect effect mediated by land titles accounts for 80% of the total impact.
- Sequential g-estimation following Acharya et al. (2016) corroborates these results and shows that two-thirds of the total effect of mobile internet on the probability of accessing a bank loan is not mediated through land titles (see appendix Table A2).

### Robustness checks and alternative specifications
- Aggregation at the bank-municipality-year level (model (6)) with bank-municipality fixed effects (α_bm) and bank-year fixed effects (φ_bt) yields:
  - a one standard deviation increase in 3G coverage is associated with a 13% increase in the number of loans granted,
  - a similar increase in the number of borrowers switching from the microfinance sector,
  - a 33% increase in total loan outstanding,
  - no significant effect on average interest rates (Table B1).
- Sample composition:
  - Original dataset: 190,138 individual borrowers.
  - Dropping borrowers with missing age (balanced panel creation) does not affect coefficients; full-sample estimates in Table B2 retain sign, magnitude, and significance.
- Alternative dependent variable definitions:
  - Using a staggered measure that takes value 1 from the first year the borrower has an outstanding loan yields results coherent with Table 3; the coefficient magnitude for bank loans is higher, while that for MFI loans is lower (Table B3).
- Alternative predictors and weightings:
  - Replacing share of 3G coverage with a dummy (1 from first year municipality reached by mobile internet) yields an increase in probability of having a bank loan of about 10 percentage points and a similar decrease for MFI loans (Table B4).
  - Weighting 3G coverage by population density and total population produces larger magnitudes, consistent with Table 3 (Table B5).
- Instrumentation strategies:
  - Main 2SLS uses two instruments for one endogenous variable to allow heterogeneous treatment effects.
  - Tests following Mogstad et al. (2021):
    - Null hypothesis of negative weights rejected at conventional levels (p = 0.000).
    - Test of null of positive weights not rejected (p-value = 1.0).
  - Instrumenting with a single instrument (lightning strikes or incidental coverage separately) yields consistent results (Table B7).
  - Replacing instruments with predetermined geographical features (altitude; percentage covered by wood, water, and sand) confirms Table 3 with larger magnitudes and F-statistics above 100 (Table B8).
  - Reduced-form regressions using lightning strikes, incidental coverage, altitude, wood, water, and sand are consistent with main specifications (Table B9).
- Controls for confounders:
  - Municipality-level controls: 2G coverage; total population; urban population; employed population; number of total bank branches; dummy for presence of a primary road. For variables from the 2012 Census, interactions with post-2010 are used.
  - Augmented specifications retain signs and statistical significance; magnitudes in some columns are slightly affected (Table B10).
- Policy program checks:
  - Inclusion of a dummy for presence of a U-SACCO (Umurenge SACCOs program, effective 2011) does not affect coefficients, indicating orthogonality between mobile internet effects and U-SACCO expansion (Table B11).
- Fixed broadband falsification and collateral channel:
  - National fiber optic backbone rollout: begun 2008, completed end-2010, with over 3,000 kilometers of fiber distributed to all 30 districts.
  - Presence of fixed broadband at municipality level included as a control (constructed from AfTerFibre Maps and Liquid Telecom private network); results align with Table 3 (Table B12).
  - Instrumenting fixed broadband using presence of post-colonial surfaced roads (map of surfaced roads in Rwanda in 1975 digitized) provides robust evidence in favor of the collateral channel (Table B13; Figure A4).

### Implications and overarching findings
- Expanded 3G coverage increases access to land titles, which in turn enables individuals to use land titles as collateral to access bank credit.
- Expanded 3G coverage is associated with:
  - a higher probability of obtaining a bank loan,
  - a decrease in the likelihood of securing a microfinance loan,
  - no significant change in overall access to finance,
  - a transition of borrowers from microfinance institutions to commercial banks, facilitated by digitally-empowered land collateral.
- The interplay between mobile internet, collateral, and access to formal credit underscores the importance of considering links across different markets when evaluating the effects of a general-purpose technology such as mobile internet.

*Source: Excerpt from the referenced PDF chapter.*

### References

### wpiea2024070-print-pdf - References

### References (bibliographic coverage)
- Literature spans land tenure and land registry reforms, microfinance and credit market information, technology and banking, mobile internet and development, and political effects of media and connectivity.
- Notable repeated topics and authors:
  - Land tenure, land registry sustainability, and land transactions: Abbott and Mugisha(2015); Ali, Deininger, and Duponchel(2017); Deininger and Goyal(2012); Manysheva(2021); Schreiber(2017).
  - Mobile internet, broadband, and development outcomes: Bahia et al.(2023); Björkegren(2019); Caldarola et al.(2023); Hjort and Poulsen(2019); Hjort and Tian(2024); GSMA(2015); UN-OHRLLS(2017).
  - Financial inclusion, microcredit, and bank-MFI dynamics: Banerjee et al.(2015); Agarwal et al.(2023); Kaboski and Townsend(2012); Beck and Cull(2013); de Mel et al.(2009).
  - Information, technology, and financial markets: Fuster et al.(2019); Gao and Huang(2020); Pierri and Timmer(2022); D’Andrea and Limodio(2024).
  - Econometrics and program evaluation: Imbens and Angrist(1994); Imbens and Wooldridge(2009); Mogstad, Torgovitsky, and Walters(2021).
- Empirical contexts include Rwanda (multiple sources), India, Brazil, El Salvador, and broader cross-country studies.

### Figures (descriptions and data sources)
- Figure (1): 3G coverage in Rwanda
  - Plots diffusion of 3G mobile internet by municipality (percentage of coverage).
  - Left panel: 2011 (first year of 3G introduction). Right panel: 2015 (last year with data).
  - Source: GSMA.
- Figure (2): Cross-sectional variation — Lightning frequency and Incidental coverage
  - Left panel: average frequency of lightning strikes by municipality (yearly average during 1998-2013).
  - Right panel: incidental coverage (fictitious coverage if towers built along full network of power lines). Source: GHRC and Björkegren (2019).
- Figure (3): 3G coverage, lightning frequency, and incidental coverage
  - Plots Std 3G mobile coverage over years (2008–2016) by above/below median lightning frequency and incidental coverage; used as visual evidence for instrument relevance.
- Figure (4): Registration time for land titles
  - Average number of days to register a land transaction between 2004 and 2019.
  - Red solid line: Rwanda. Green dashed line: average across Sub-Saharan Africa countries.
  - Vertical lines mark years of the LT reform. Source: World Bank Doing Business.

### Tables — key statistics, estimates, and diagnostics

- Table (1): Summary statistics (selected exact figures)
  - Panel A: Access to bank credit (n = 909165)
    - Probability of Any Loan: 0.280 mean, 0.45 sd, 0.00 min, 0.00 50, 1.00 max
    - Probability of Bank Loan: 0.140 mean, 0.34 sd, 0.00 min, 0.00 50, 1.00 max
    - Probability of MFI Loan: 0.160 mean, 0.37 sd, 0.00 min, 0.00 50, 1.00 max
    - Switching borrowers: 0.090 mean, 0.29 sd, 0.00 min, 0.00 50, 1.00 max
  - Panel B: Bank loan characteristics
    - Outstanding loan (n = 124,806): mean 2011.88, sd 3665.80, 50 693.11, min 16.67, max 27859.83 (units: thousands of Rwandan francs)
    - Principal loan (n = 125,507): mean 2820.34, sd 706.01, 50 1062.00, min 30.00, max 34000
    - Interest rate (n = 124,388): mean 19.20, sd 5.73, 50 19.00, min 6.14, max 70.10 (percentage terms)
    - Collateral (n = 128,639): mean 0.12, sd 0.32, 50 0.00, min 0.00, max 1.00 (dummy)
  - Panel C: 3G coverage (n = 26960)
    - 3G coverage: mean 0.120, sd 0.280, 50 0.00, min 0.00, max 1.00
    - Std 3G coverage: mean 0.031, sd 1.02, 50 -0.40, min -0.40, max 3.24
  - Panel D: Instrumental variables (n = 26960)
    - Lightning frequency: mean 0.040, sd 0.020, 50 0.03, min 0.02, max 0.10
    - Std Lightning: mean 0.030, sd 0.830, 50 0.00, min -1.24, max 4.10
    - Incidental coverage: mean 0.661, sd 12.20, 50 9.54, min 0.00, max 135.43
    - Std Incidental: mean 0.010, sd 0.830, 50 0.00, min -0.91, max 10.81
  - Notes: Std Lightning and Std Incidental are interactions with a post-2010 dummy; standardizations implemented on all Rwandan municipalities.

- Table (2): First stage regressions (dependent variable: Std 3G coverage)
  - Std Lightning coefficients:
    - Column (1): -0.156*** (0.048)
    - Column (2): -0.141*** (0.046)
    - Column (3): -0.341*** (0.079)
    - Column (4): -0.330*** (0.079)
  - Std Incidental coefficients:
    - Column (1): 0.250*** (0.052)
    - Column (2): 0.240*** (0.050)
    - Column (3): 0.231*** (0.067)
    - Column (4): 0.207*** (0.064)
  - Fixed effects: Municipality FE (Columns 1–3 as noted), Borrower FE (Columns 4–6 where specified), Year FE (all columns).
  - F-statistics: 10.64, 23.17, 17.74, 18.49, 11.81, 16.22 (by column)
  - Observations reported across columns: 26962 / 6962 / 2696 / 909165 / 909165 / 909165 (as listed)
  - Adj. R sq.: 0.5400, 0.5510, 0.5570, 0.6070, 0.5930, 0.613

- Table (3): Mobile Internet and the Probability of having a loan
  - OLS estimates (Std 3G coverage):
    - Any Loan (Col 1): 0.002 (0.002)
    - Bank Loan (Col 2): 0.018*** (0.003)
    - MFI Loan (Col 3): -0.017*** (0.003)
  - IV (2SLS) estimates (Std 3G coverage):
    - Any Loan (Col 4): -0.002 (0.006)
    - Bank Loan (Col 5): 0.034*** (0.012)
    - MFI Loan (Col 6): -0.036*** (0.012)
  - SW F-statistic: 16.22 (reported)
  - Observations: 909165 (all columns)
  - Adj. R sq. (reported for OLS columns): 0.4800 (Any Loan), 0.3740 (Bank Loan), 0.370 (MFI Loan)
  - Mean Dep. Var. and S.D. Dep. Var. reported precisely per column.

- Table (4): Mobile Internet and Transition to banks (switching behavior)
  - OLS estimates (Std 3G coverage):
    - Probability of Switching from MFI to Bank (Col 1): 0.007*** (0.002)
    - Probability of Switching from Bank to MFI (Col 3): -0.006*** (0.001)
  - IV estimates (Std 3G coverage):
    - Probability of Switching from MFI to Bank (Col 2): 0.011*** (0.003)
    - Probability of Switching from Bank to MFI (Col 4): -0.011** (0.005)
  - SW F-statistic: 16.47 and 13.89 (as reported)
  - Observations:
    - Switching from MFI to Bank: 514372 (Cols 1–2)
    - Bank to MFI: 380974 (Cols 3–4)

- Table (5): Heterogeneity — Socioeconomic and Borrowers’ characteristics (Probability of Bank Loan, IV estimates)
  - Panel A (municipality splits)
    - Rural (Col 1): Std 3G coverage 0.009 (0.021)
    - Urban (Col 2): Std 3G coverage 0.032* (0.017)
    - Poorer (Col 3): Std 3G coverage 0.003 (0.032)
    - Wealthier (Col 4): Std 3G coverage 0.044*** (0.011)
    - No branches (Col 5): Std 3G coverage 0.024 (0.017)
    - Branches (Col 6): Std 3G coverage 0.034** (0.016)
  - Panel B (borrower splits)
    - Female (Col 1): Std 3G coverage 0.029** (0.012)
    - Male (Col 2): Std 3G coverage 0.037*** (0.013)
    - Single (Col 3): Std 3G coverage 0.050*** (0.017)
    - Married (Col 4): Std 3G coverage 0.030** (0.012)
    - Young (Col 5): Std 3G coverage 0.037*** (0.012)
    - Adult (Col 6): Std 3G coverage 0.030** (0.014)
  - SW F-statistics vary by column (examples: 10.51, 8.47, 9.88, 9.31, 12.32, 6.64 in Panel A; 14.17, 16.89, 12.84, 16.46, 16.64, 15.62 in Panel B).
  - Observations reported per column (examples include 482259, 426906, 338929, 570236, 397445, 511720 in Panel A; 324226, 584939, 106749, 802416, 466877, 442288 in Panel B).

- Table (6): Mobile Internet and Bank loan characteristics
  - Dependent variables: Volume of Loans (ln total outstanding loans) and Lending Rate (average interest rate).
  - IV estimates (Std 3G coverage):
    - Column (1) Volume of Lending: 0.063** (0.025)
    - Column (2) Lending Rate: 0.154 (0.154)
    - Column (3) Volume of Lending (with Switcher interaction): 0.032 (0.032)
    - Column (4) Lending Rate (with Switcher interaction): -0.077 (0.207)
  - Interaction term Std 3G coverage × Switcher:
    - Coefficient: 0.159*** (0.054) on Volume of Lending
    - Coefficient: -2.235*** (0.660) on Lending Rate
  - F-test (p-value) for interaction: 0.001 and 0.001
  - SW F-statistics: 22.44, 21.64, 18.96, 19.00 (by column)
  - Observations and means:
    - Obs. examples: 104766, 104423, 110391, 79081
    - Mean Dep. Var. examples: 6.750, 18.51, 6.609, 20.71
    - S.D. Dep. Var. examples: 1.4215, 5.4721, 1.2878, 8.603

- Table (7): Mobile Internet and Land transactions (municipality-level)
  - Dependent variables: 50% Benchmark (dummy that transactions reached 50% of total in first year of reform) and Number of Transactions (inverse hyperbolic sine of cumulative transactions scaled by parcels).
  - OLS estimates:
    - Std 3G coverage (50% Benchmark, Col 1): 0.173*** (0.020)
    - Std 3G coverage (Number of Transactions, Col 2): 0.047*** (0.007)
  - IV estimates:
    - Std 3G coverage (50% Benchmark, Col 3): 0.147* (0.080)
    - Std 3G coverage (Number of Transactions, Col 4): 0.128*** (0.036)
  - SW F-statistic: 19.19 (reported)
  - Observations: 2022 (all columns)
  - Adj. R sq.: 0.5770 (Col 1), 0.804 (Col 2)
  - Mean Dep. Var.: 0.122 (50% Benchmark), 0.229 (Number of Transactions)
  - S.D. Dep. Var.: 0.3270, 0.3160

- Notes across tables:
  - Main predictor throughout: Std3G coverage (standardized 3G mobile coverage, municipality m, time t).
  - Instruments: Z_light (standardized yearly average lightning frequency interacted with post-2010 dummy) and Z_incid (standardized incidental coverage interacted with post-2010 dummy), and interactions with Switcher where applicable.
  - Fixed effects commonly include municipality, borrower, bank-year, and year as specified per regression.
  - Standard errors are clustered at the municipality level. Significance indicators: ***, **, * denote 1%, 5% and 10% levels, respectively.

*Italic: References, figure notes, and table notes are reproduced exactly as provided in the source content.*

### 2010.  Columns 1 and 2 refer to OLS estimates.  Columns 3 and 4 refer to 2SLS estimates.  SW F-statistics reports

### wpiea2024070-print-pdf - 2010.  Columns 1 and 2 refer to OLS estimates.  Columns 3 and 4 refer to 2SLS estimates.  SW F-statistics reports

### Table (8): Mobile Internet, Land mortgaged and Bank collateral
- Dependent variables:
  - Share(Land Mortgaged): the share of land mortgaged in municipality m.
  - Probability(Bank Collateral): dummy = 1 if borrower i in municipality m has an outstanding loan in year t backed by collateral.
  - Probability(Bank Mortgage): dummy = 1 if borrower i in municipality m has mortgage with a bank in year t.
- Main predictor:
  - Std 3G coverage: the standardized measure of 3G mobile coverage in municipality m at time t, instrumented by:
    - Z_light_mt: the standardized yearly average frequency of lightning in municipality m, interacted with a dummy post-2010.
    - Z_incid_mt: the standardized average of incidental coverage in municipality m, interacted with a dummy post-2010.
- Estimation and fixed effects:
  - Columns 1–2: OLS.
  - Columns 3–4: 2SLS (IV).
  - Borrower FE: No in columns 1–2; Yes in columns 3–6.
  - Year FE: No in columns 1–2; Yes in columns 3–6.
- Reported coefficient estimates on Std 3G coverage (with clustered standard errors in parentheses):
  - Column (1): 0.024*** (0.003)
  - Column (2): 0.017*** (0.005)
  - Column (3): 0.009*** (0.001)
  - Column (4): 0.003*** (0.001)
  - Column (5): 0.006** (0.003)
  - Column (6): 0.008*** (0.002)
- SW F-statistics (Sanderson-Windmeijer multivariate F-statistics): 18.52 16.22 16.22
- Sample size and fit statistics:
  - Obs.: 337 337 909165 909165 909165 909165 (as reported)
  - Adj. R sq.: 0.528 0.499 0.512
- Dependent variable moments (as reported):
  - Mean Dep. Var.: 0.014 0.014 0.015 0.016 0.015 0.016
  - S.D. Dep. Var.: 0.034 0.034 0.120 0.125 0.120 0.125
- Notes on inference:
  - Fixed effects are at the municipality and year levels (columns 3 to 6 specify borrower and year level in notes).
  - Standard errors are clustered at the municipality or district level as indicated.
  - ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

### Table (9): Sobel-Goodman mediation test
- Outcome variables:
  - Probability(Bank Loan): dummy = 1 if borrower i in municipality m has an outstanding loan with a bank in year t.
  - Probability(Bank Collateral): dummy = 1 if borrower i in municipality m has an outstanding loan in year t backed by collateral.
- Mediator:
  - 50%Benchmark: dummy = 1 if the number of land transactions in municipality m has reached 50% of the total in the first year of the reform.
- Reported mediation test statistics and effects (all values reported exactly):
  - Sobel: 0.013** (Probability of Bank Loan); 0.005** (Bank Collateral)
  - Aroian: 0.013**; 0.005**
  - Goodman: 0.013**; 0.005**
  - Indirect Effect (Collateral): 0.013**; 0.005**
  - Direct Effect (Mobile Internet): 0.021**; 0.001
  - Total Effect: 0.034**; 0.006
  - Proportion Indirect: 37% (for Bank Loan); 80% (for Bank Collateral)
- Interpretation provided:
  - Table reports proportion of the total effect of mobile Internet on the dependent variable that is mediated through land titles.
  - ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

### Additional figures and descriptive evidence
- Figure (A1): Dynamic effect of mobile internet on access to bank credit
  - Design: staggered DiD event study.
  - Treatment: borrowers in municipalities covered by 3G technology.
  - Control: borrowers in municipalities not reached by 3G technology.
  - Year 0: first year municipality is reached by mobile internet.
  - Y-axis: Prob.(bank loan), the probability borrower has an outstanding loan with a commercial bank.
  - Visual elements: blue solid line for coefficients with 90% confidence intervals.
- Figure (A2): Distribution of interest rates for borrowers switching from MFIs to banks
  - Plots density of interest rates for switchers.
  - Blue solid line: density before switching (when borrower is a client of the MFI).
  - Red solid line: density after the switch.
- Figure (A3): Share of mortgaged land parcels
  - Plots share of mortgaged land parcels by municipality in Rwanda.
  - Source: Land Management and Use Authority.
- Figure (A4): Distribution of fiber-optic cables and post-colonial roads
  - Left panel: fiber-optic lines as in 2015 (public fiber deployed in 2010; private fiber complemented main lines from 2014).
  - Right panel: surfaced roads as in 1975, from the Perry-Castañeda Library Map Collection.

### Table (A1): Balance table — Municipalities by quartiles of lightning and incidental coverage
- Purpose: reports balance on average municipalities’ characteristics in 2010 by quartiles of the instruments.
- Panel A: Lightning Strikes (Z_light)
  - Quartile thresholds used: 25th, 50th, 75th percentiles.
  - Example reported values (Below / Above / Normalized difference):
    - 2G coverage: .93 .928 (.01) ; .937 .92 (.08) ; .924 .942 (-.09) — N. mun.: 337
    - Mean altitude: 1.499 1.839 (-1.06) ; 1.582 1.922 (-.91) ; 1.673 1.988 (-.69) — N. mun.: 336/337 depending on row
    - ln Population: 10.175 10.03 (.31) ; 10.117 10.016 (.22) ; 10.057 10.095 (-.09) — N. mun.: 337
    - ln Urban: 3.583 2.644 (.15) ; 3.622 2.138 (.24) ; 2.736 3.298 (-.09) — N. mun.: 337
    - Night lights: .404 .700 (-.12) ; .763 .49 (.09) ; .674 .484 (.07) — N. mun.: 337
    - Bank branches: .762 .743 (.01) ; .863 .633 (.10) ; .75 .741 (0.00) — N. mun.: 337
- Panel B: Incidental Coverage (Z_incid)
  - Quartile thresholds used: 25th, 50th, 75th percentiles.
  - Example reported values (Above / Below / Normalized difference):
    - 2G coverage: .956 .845 (.49) ; .969 .888 (.41) ; .974 .913 (.36) — N. mun.: 337
    - Mean altitude: 1.734 1.806 (-.16) ; 1.693 1.810 (-.27) ; 1.669 1.779 (-.23) — N. mun.: 336
    - ln Urban: 3.481 1.063 (.44) ; 4.237 1.527 (.45) ; 3.191 2.774 (.07) — N. mun.: 337
    - Night lights: .77 .193 (.23) ; 1.006 .249 (.27) ; .774 .577 (.06) — N. mun.: 337
    - Bank branches: .866 .393 (.25) ; 1.054 .444 (.28) ; .583 .802 (.11) — N. mun.: 337
- Notes on balance table:
  - Columns 1–3 use the 25th percentile threshold; Columns 4–6 use the 50th percentile; Columns 7–9 use the 75th percentile.
  - Column 10 reports the number of municipalities.
  - Normalized difference computed as proposed in Imbens and Wooldridge (2009): difference in averages by treatment status scaled by square root of sum of variances; scale-free measure of distribution difference.

*Content extracted from wpiea2024070-print-pdf - 2010.  Columns 1 and 2 refer to OLS estimates.  Columns 3 and 4 refer to 2SLS estimates.  SW F-statistics reports*

### 0.25  are  considered  sensitive  for  linear  regressions.   The  list  of  covariates  under  evaluation  is  the  fol

### wpiea2024070-print-pdf

### Covariates under evaluation
- Coverage of 2G
- Mean altitude
- Percentage of wood, water, and sand
- Indicator of the presence of a main line power/hospital
- Log of the population and the urban population
- Unemployment rate
- Indicator of night lights and poverty (computed with respect to the average value in the district)
- Measure of conflict
- Number of bank branches

### Sequential g-estimation (Table A2) — key estimates and sample metrics
- Std 3G coverage coefficients:
  - Column (1) Probability of Bank Loan (OLS): 0.008*** (0.003)
  - Column (2) Bank Loan (IV): 0.020* (0.011)
  - Column (3) Total Effect (IV benchmark): 0.034*** (0.012)
- Estimation methods: OLS and 2SLS
- Fixed effects: Borrower FE Yes; Year FE Yes
- SW F-statistic: 16.22 16.22
- Obs.: 909165 909165 909165
- Adj. R sq.: 0.001
- Mean Dep. Var.: 0.136 0.136 0.136
- S.D. Dep. Var.: 0.343 0.343 0.343

### Mobile internet and aggregate bank characteristics (Table B1)
- Std 3G coverage coefficients:
  - Number of Loans (col 1): 0.130** (0.065)
  - Number of Switchers (col 2): 0.118*** (0.034)
  - Volume of Loans (col 3): 0.285*** (0.091)
  - Lending Rate (col 4): 0.279 (0.287)
- Estimation: IV (2SLS)
- Fixed effects: Bank-municipality FE Yes; Bank-year FE Yes
- SW F-statistic: 15.97 15.97 15.62 15.70
- Obs.: 126951 126951 2459 12467
- Mean Dep. Var.: 1.307 0.171 8.650 18.75
- S.D. Dep. Var.: 1.301 0.420 1.696 5.375

### Mobile internet and access to credit — All borrowers (Table B2)
- Std 3G coverage:
  - Probability of Bank Loan (col 1): 0.033*** (0.012)
  - MFI Loan (col 2): -0.035*** (0.013)
- Estimation: IV
- Fixed effects: Borrower FE Yes; Year FE Yes
- SW F-statistic: 15.76 15.76
- Obs.: 1.268e+06 1.268e+06
- Mean Dep. Var.: 0.157 0.190
- S.D. Dep. Var.: 0.363 0.392

### Mobile internet and access to credit — First relationship (Table B3)
- Std 3G coverage:
  - Probability of Bank Loan (col 1): 0.048*** (0.015)
  - MFI Loan (col 2): -0.025* (0.014)
- Estimation: IV
- Fixed effects: Borrower FE Yes; Year FE Yes
- SW F-statistic: 16.22 16.22
- Obs.: 909165 909165
- Mean Dep. Var.: 0.163 0.189
- S.D. Dep. Var.: 0.369 0.392

### Two-way fixed effects and dummy 3G (Table B4)
- Dummy 3G:
  - Probability of Bank Loan (col 1): 0.094*** (0.036)
  - MFI Loan (col 2): -0.110*** (0.035)
- Estimation: IV
- Fixed effects: Borrower FE Yes; Year FE Yes
- SW F-statistic: 19.17 19.17
- Obs.: 909165 909165
- Mean Dep. Var.: 0.136 0.164
- S.D. Dep. Var.: 0.343 0.370

### Weighted 3G coverage (Table B5)
- Std 3G coverage (weighted):
  - Probability of Bank Loan (col 1): 0.051** (0.022)
  - MFI Loan (col 2): -0.047** (0.020)
  - Bank Loan (col 3): 0.049*** (0.017)
  - MFI Loan (col 4): -0.048*** (0.017)
- Weighting: population density (cols 1–2) and total population (cols 3–4)
- Estimation: IV
- Fixed effects: Borrower FE Yes; Year FE Yes
- SW F-statistic: 9.669 9.669 11.63 11.63
- Obs.: 905574 905574 905574 905574
- Mean Dep. Var.: 0.137 0.163 0.137 0.163
- S.D. Dep. Var.: 0.343 0.370 0.343 0.370

### Alternative instruments and single-instrument specifications (Tables B6–B7)
- Table B6 Std 3G coverage (various instrument versions):
  - Cols 1–4 coefficients: 0.056*** (0.017), -0.086*** (0.022), 0.047*** (0.015), -0.070*** (0.018)
  - SW F-statistic: 11.55 11.55 13.06 13.06
  - Obs.: 909165 repeated
  - Mean Dep. Var.: 0.136 0.164 0.136 0.164
  - S.D. Dep. Var.: 0.343 0.370 0.343 0.370
- Table B7 Std 3G coverage (single instruments):
  - Cols 1–4 coefficients: 0.036** (0.015), -0.035** (0.015), 0.018 (0.017), -0.037** (0.018)
  - SW F-statistic: 18.49 18.49 11.81 11.81
  - Obs.: 909165 repeated
  - Mean Dep. Var.: 0.136 0.164 0.136 0.164
  - S.D. Dep. Var.: 0.343 0.370 0.343 0.370

### Geographical instruments (Table B8)
- Std 3G coverage:
  - Probability of Bank Loan (col 1): 0.048*** (0.009)
  - MFI Loan (col 2): -0.049*** (0.009)
- Instruments: two main instruments plus mean altitude and percentage territory covered by wood, water, and sand, all interacted with a dummy post-2010
- Estimation: IV
- Fixed effects: Borrower FE Yes; Year FE Yes
- SW F-statistic: 861.54 861.54
- Obs.: 906990 906990
- Mean Dep. Var.: 0.136 0.163
- S.D. Dep. Var.: 0.343 0.369

### Reduced-form estimates for instruments and geographical covariates (Table B9)
- Std Lightning:
  - Bank Loan: -0.013** (0.006)
  - MFI Loan: 0.012** (0.006)
- Std Incidental:
  - Bank Loan: 0.003 (0.004)
  - MFI Loan: -0.008* (0.004)
- Altitude:
  - Bank Loan: -0.021*** (0.006)
  - MFI Loan: 0.024*** (0.006)
- Wood:
  - Bank Loan: 0.001 (0.012)
  - MFI Loan: 0.003 (0.009)
- Water:
  - Bank Loan: -0.008** (0.003)
  - MFI Loan: 0.008* (0.004)
- Sand:
  - Bank Loan: 0.014*** (0.001)
  - MFI Loan: -0.008*** (0.001)
- Estimation: OLS
- Fixed effects: Borrower FE Yes; Year FE Yes
- Obs.: 909165 909165 906920 906920 (by column)
- Adj. R sq.: 0.372 0.369 0.374 0.370
- Mean Dep. Var.: 0.136 0.163 0.136 0.163
- S.D. Dep. Var.: 0.343 0.369 0.343 0.369

### Additional controls and program controls (Tables B10–B12)
- Table B10 (Additional controls):
  - Std 3G coverage: 0.037*** (0.013) for Bank Loan; -0.028* (0.014) for MFI Loan
  - SW F-statistic: 17.34 17.34
  - Obs.: 931774 931774
  - Mean Dep. Var.: 0.155 0.183
  - S.D. Dep. Var.: 0.361 0.386
- Table B11 (Controlling for the U-SACCO program):
  - Std 3G coverage: 0.035*** (0.012) for Bank Loan; -0.037*** (0.011) for MFI Loan
  - SW F-statistic: 16.26 16.26
  - Obs.: 909165 909165
  - Mean Dep. Var.: 0.136 0.164
  - S.D. Dep. Var.: 0.343 0.370
- Table B12 (Controlling for broadband internet):
  - Std 3G coverage: 0.033*** (0.012) for Bank Loan; -0.035*** (0.011) for MFI Loan
  - SW F-statistic: 16.31 16.31
  - Obs.: 909165 909165
  - Mean Dep. Var.: 0.136 0.164
  - S.D. Dep. Var.: 0.343 0.370

### Broadband internet and land transactions (Table B13)
- Std 3G coverage:
  - 50% Benchmark (col 1): 0.162** (0.082)
  - Number of Transactions (col 2): 0.013*** (0.037)
- Fiber:
  - 50% Benchmark (col 1): 0.002 (0.063)
  - Number of Transactions (col 2): -0.024 (0.028)
- Estimation: IV (2SLS) at municipality level
- Fixed effects: Municipality FE Yes; Year FE Yes
- SW F-statistic: 19.43 19.43
- Alternative SW F-statistic reported: 50.83 50.83
- Obs.: 2022 2022
- Mean Dep. Var.: 0.121 0.229
- S.D. Dep. Var.: 0.327 0.316

*Italic: Source: wpiea2024070-print-pdf (canonical URL provided separately).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024070-print-pdf.pdf_