## wpiea2021214-print-pdf

## Source details

**Canonical URL:** [wpiea2021214-print-pdf](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021214-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021214-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021214-print-pdf.pdf.json)

---

### Model setup: agents, heterogeneity, and project technology
- Firms: unit mass normalized to one; share of young firms is y∈(0,1). Firm age and collateral are independent.
- Collateral: drawn from continuous distribution G(C). Market price of collateral at date 1 is P.
- Project payoff: yields x > 1 if successful and 0 if unsuccessful.
- Project quality (private to firm): good (G) with success probability p_G and bad (B) with success probability p_B, where 0 < p_B < p_G < 1 and only good projects have positive NPV: p_B x < 1 < p_G x.
- Share of good projects at date 0 is q > 0, independent of bank or firm characteristics.
- Adverse selection assumption (severe): [q p_G + (1−q) p_B] x < 1, so banks screen all borrowers in equilibrium and as a result all loans granted are to good firms.

### Banks, IT adoption, and screening tools
- Banks: unit mass of banks, each endowed with one unit of investment good at date 0.
- IT adoption: exogenous fraction h∈(0,1) of banks are high-IT; remainder are low-IT.
- Screening tools:
  - Information acquisition: bank pays fixed cost F to learn project type. Fixed cost is lower for old firms than young firms: F_O < F_Y.
  - Collateral-based screening: bank asks for collateral at date 0, repossessed and sold at date 1 if default. Cost of screening via collateral is lower for high-IT banks than low-IT banks: v_HighIT < v_LowIT.
- Lending volume: maximizes joint surplus; banks receive fraction θ∈(0,1) of surplus. Banks and firms are randomly matched; startups do not apply to multiple banks (no multi-application).

### Assumed ranking of screening costs (baseline)
- v_HighIT < F_O < p_G x − 1 < min{F_Y, v_LowIT}.

### Incentive compatibility and minimum collateral
- Only good firms may receive credit in equilibrium (fraction q).
- Young firms receive credit only when matched with a high-IT bank (fraction h of banks) and if collateral C > C_min, where C_min solves:
  - p_B (x − r) ≡ (1 − p_B) P C_min,
  - r is the bank’s lending rate.
- Interpretation: left-hand side is benefit to a bad firm from pretending to be good; right-hand side is cost from losing collateral value upon failure. C_min depends on P: C_min = C_min(P).

### Equilibrium screening for old firms
- Old firms always receive credit.
  - Matched to high-IT bank: lending is backed by collateral if C ≥ C_min, otherwise bank acquires information (pays F_O).
  - Matched to low-IT bank: screening via information acquisition is used.

### Proposition and main comparative-statics (share of lending to young firms)
- Proposition 1: share of lending to young firms is
  - s_Y ≡ y h [1 − G(C_min)] / [1 − y + y h [1 − G(C_min)]].
- Predictions (comparative statics):
  - Prediction 1: ds_Y / dh > 0. A higher share of high-IT banks increases the share of lending to young firms.
  - Prediction 2: ds_Y / dP > 0. A higher collateral value increases the share of lending to young firms.
  - Prediction 3: d^2 s_Y / (dh dP) > 0. A higher collateral value increases the share of lending to young firms more when the share of high-IT banks is higher.
  - Prediction 4: Bank IT adoption does not affect the quality (default rate) of firms receiving funding in equilibrium (all funded projects are good).
- Additional model implication: young firms use more collateral than old firms in equilibrium.

### Extensions and additional predictions
- Extension 1: Recourse versus non-recourse states
  - Fraction i∈(0,1) of firm owners generate external income I ≥ P C_min; banks in recourse states can access I upon default; in non-recourse states they cannot.
  - Collateral and recourse are substitutes for deterring bad firms.
  - Prediction 5: ds_Y / dh is larger in non-recourse than in recourse states:
    - ds_Y / dh | Non−recourse > ds_Y / dh | Recourse.
  - Similarly, the interaction effect with collateral is stronger in non-recourse states:
    - d^2 s_Y / (dh dP) | Non−recourse > d^2 s_Y / (dh dP) | Recourse.
- Extension 2: Distance (soft information and proximity)
  - Allow low-IT banks to screen some young firms that are close by by assuming:
    - F_close^Y < p_G x − 1 < F_distant^Y < v_LowIT.
  - Prediction 6: Distance matters less for high-IT banks than for low-IT banks; high-IT bank lending is relatively insensitive to distance shocks, low-IT bank lending is more sensitive and favors close young firms.

### Empirical implementation: data construction and variables
- Sample period: 1999–2007 (main analysis), chosen to avoid post-crisis regulatory changes.
- Bank IT adoption measure:
  - Source: establishment-level survey on personal computers per employee by CiTBDs Aberdeen for years 1999, 2003, 2004, and 2006; focus on banking sector establishments (SIC2), 143,607 establishment-year observations.
  - Construct bank-level IT measure ̃IT_b by regressing PCs/Emp on a bank fixed effect and controls (establishment type, county, year, log employees) for years 1999, 2003, 2004, 2006.
  - County-level IT exposure IT_c constructed using branch distribution in 1994:
    - IT_c = Σ_{b=1}^N ̃IT_b * No.Branches_{b,c} / No.Branches_c,
    - standardized to mean zero and standard deviation one.
  - Correlations: IT budget (2016) correlates with PCs per employee at 0.65; cross-sectional R^2 of PCs per Employee on per capita IT budget is 44%.
- Young firm measure:
  - Source: Quarterly Workforce Indicators (QWI), employment at county-two-digit NAICS industry-year level by firm age brackets.
  - Define young firms as age 0–1. Use 4th quarter values; employment of young firms is a flow (jobs created by new firms).
  - Baseline normalization: job creation of young firms scaled by total employment in same county-industry cell.
- Industry-level home equity dependence:
  - Source: 2007 Public Use Survey of Business Owners (SBO); compute fraction of young firms reporting use of home equity or personal assets to start/expand, aggregated by 2-digit industry.
- County controls: log total population, share black population, share population older than 65, unemployment rate, house price growth (FHFA HPI), log per capita income.
- Bank balance-sheet data: FDIC SDI (second quarter each year) for assets, Tier 1, non-interest & total income, investment securities, efficiency ratio, non-performing loans, ROA, deposits.
- Small business lending response: CRA data on loan originations below $1 million by financial institutions with > $1 billion assets, aggregated to BHC-county level. Year-to-year change normalized by midpoint:
  - ∆CRA_{b,c,t} = (CRA_{b,c,t} − CRA_{b,c,t−1}) / ((CRA_{b,c,t} + CRA_{b,c,t−1})/2),
  - bounds growth in [−2, 2].

### Main empirical tests and findings (tests of model predictions)
- Empirical specification for Prediction 1 (cross-sectional at county-industry level):
  - startups_{c,i} = β_1 IT_exposure_{c,99} + β_2 constraint_i + β_3 IT_exposure_{c,99} × constraint_i + controls_{c,99} + θ_c + φ_i + ε_{c,i},
  - startups_{c,i} is employment share of firms age 0–1 out of total employment, averaged 1999–2007.
- Reduced-form evidence:
  - Nonparametric plot (Figure 2) shows positive relation between IT exposure and startup employment.
  - Table 3 regressions:
    - Column (1): positive relation between county IT exposure and startup share.
    - Column (2): coefficient stable after county-level controls.
    - Column (3): including NAICS2 industry fixed effects leaves coefficient statistically and economically similar; one standard deviation higher IT exposure associated with 0.38 pp increase in share of young firm employment (4% of mean 9.3%).
    - Column (4): interaction with industry external-finance dependence positive and significant; a one standard deviation higher IT exposure associated with 1 pp increase in young firm employment share in industries dependent on external finance (11% of mean).
    - Column (5): adding county fixed effects yields similar results (interaction estimate changes by 0.02 pp).
  - Results interpreted as consistent with Prediction 1 and a lending channel: IT exposure matters more in industries that depend on external finance.
- Robustness exercises (Table A1 highlights):
  - Alternative IT exposure constructions (unweighted average, deposits-weighted) produce consistent results.
  - Excluding startups in financial and education industries; excluding Wyoming (high IT exposure); including state fixed effects; alternative normalization; results robust.
  - Stronger relationship within tradable industries (manufacturing and mining/extraction aggregated to 2-digit) — suggests not driven by local demand and consistent with financing channel.
  - Excluding top VC counties or seven high-VC states yields similar results — not driven by VC concentration.
  - Coverage concerns: average survey coverage is 13.6% (sd 8.4%); dropping bottom coverage quartile or including coverage as control does not change results.
- Instrumental variables approach:
  - Two-stage IV using predicted bank deposits from gravity model adjusted by staggered interstate banking deregulation to form ̂IT_c:
    - ̂IT_c = Σ_{b=1}^N ̃IT_b * ̂Deposits_{b,c} / Deposits_c.
  - IV results (Table 4): first stage shows positive correlation between actual and predicted IT exposure; reduced form positive; second stage IV impact qualitatively similar and larger in magnitude than OLS. Difference between OLS and IV estimates not statistically different from zero.
- Changes over time (difference approach):
  - Define ∆IT_c = Σ_{b=1}^N ∆̃IT_b * No.Branches_{b,c} / No.Branches_c, where ∆̃IT_b is increase in bank IT adoption between 1999 and 2006.
  - Counties more exposed to increases in IT experienced less negative decreases in startup rates (Figure 3); formal regressions in Table A2 confirm positive correlation between ∆IT and changes in startup rates.

### 4.2 IT, house prices, and entrepreneurship (Predictions 2 & 3) — theoretical predictions
- Prediction 2: d sY / dP > 0 : a higher collateral value increases the share of lending to young firms.
- Prediction 3: d2 sY / dh dP > 0 : a higher collateral value increases the share of lending to young firms more when the share of high-IT banks is higher.
- Collateral channel emphasized: rising real estate prices increase collateral values, mitigating informational frictions and relaxing borrowing constraints for small and young firms.

### 4.2 IT, house prices, and entrepreneurship — empirical strategy
- Estimated panel regression at county-industry-year level (1999–2007):
  - startups_c,i,t = γ1 IT exposure_c,99 + γ2 ∆HPI_c,t + γ3 IT exposure_c,99 × ∆HPI_c,t + controls_c,t−1 + θ_c,i + τ_t + ε_c,i,t. (Equation 13)
- Dependent variable: employment share of firms age 0–1 out of total employment in county (c) and 2-digit industry (i) in year (t).
- IT exposure_c,99: counties’ IT exposure as of 1999, standardized to mean zero and a standard deviation of one.
- ∆HPI_c,t: yearly county-level growth in house prices.
- Controls (lagged one period unless county fixed effects included): log total population, share of population age 65 and older, share of black population, education, unemployment rate, employment shares in industries 23, 31, 44, 62, and 72, and PCs per employee in non-financial firms.
- Standard errors clustered at the county level.
- Identification: county-industry fixed effects and industry×year fixed effects used to isolate within county-industry variation and absorb the level effect of IT exposure.

### 4.2 Main empirical findings (tests of Predictions 2 & 3)
- Support for Prediction 2:
  - γ2 > 0 is implied by Prediction 1; empirical estimates show a rise in house prices is associated with an increase in local entrepreneurship conditional on year fixed effects.
  - Finding persists when controlling for county IT adoption.
- Support for Prediction 3:
  - Focus on γ3 (interaction of IT exposure and ∆HPI). Estimates find γ3 > 0, consistent with Prediction 3.
  - Interaction coefficient remains positive and similar in size when adding time-varying county controls and industry×year fixed effects.
- Robustness:
  - County-industry fixed effects absorb IT exposure level; interaction captures how house-price sensitivity of entrepreneurship varies with initial IT exposure.

### 4.2 Industry heterogeneity and complementary evidence
- Industries with lower average start-up capital or greater reliance on home equity show larger responsiveness.
- Results: Larger benefits of house price increases due to presence of high-IT banks occur predominantly in industries whose financing is more sensitive to collateral-value changes.
- Appendix A1 (bank-county-year level): high-IT banks’ small business lending reacts more to increases in local house prices.
- Summary interpretation: entrepreneurship increases when local collateral values increase, especially in counties with higher exposure to IT-intensive banks — consistent with Predictions 1–3.
- Policy relevance highlighted: IT adoption in lending decisions can spur job creation by young firms by making collateralized lending cheaper; with recent house price rises and greater IT reliance, IT adoption can spur entrepreneurship and productivity growth.

### Key numerical findings and robustness summaries
- County-level IT exposure and startup employment (Table 3): IT exposure coefficients: 0.455***, 0.397***, 0.370***, 0.373*** with standard errors (0.118), (0.098), (0.098), (0.098); Observations: 25,742; R-squared across columns: 0.003, 0.047, 0.252, 0.252, 0.354.
- IV results (Table 4): IT exposure (OLS) 0.319*** (0.109); IT exposure (second stage) 0.526*** (0.143); Gravity/RS first-stage coefficients 0.640*** (0.0667) and 0.337*** (0.0889); Observations: 19,293.
- Interaction with housing collateral (Table 5): ∆ HPI coefficients include 0.025**, 0.024**, -0.024**, -0.041***, -0.034***, -0.028**; IT exposure×∆ HPI coefficients include 0.075***, 0.070**, 0.075**, 0.271*** with reported standard errors (0.027), (0.033), (0.030), (0.086).
- Transition rates of young firms (Table 6): IT exposure coefficients approximately 0.000 (standard errors 0.000); IT exposure×ext. fin. dep coefficients around -0.001 (standard errors 0.001).
- Recourse law heterogeneity (Table 7): IT exposure coefficients: 0.305*** (0.0966), 0.471*** (0.176), 0.700*** (0.203), 0.673*** (0.204). Interaction Recourse State×IT exposure: -0.463** (0.220), -0.434** (0.220).
- CRA and home equity lending responsiveness (Tables A5–A6): IT×∆ HPI coefficients and ∆ HPI coefficients reported with exact estimates and standard errors; examples include IT×∆ HPI 0.213*** (0.066) and ∆ HPI 2.158*** (0.090).
- Descriptive statistics (Table 1, Obs = 1774): IT exposure Mean -0.001, Std. Dev. 1.235, Min -0.562, Max 0.964; log(pop) Mean 10.995, Std. Dev. 1.135; log(income pc) Mean 10.062, Std. Dev. 0.206; unemployment rate mean 4.67 (Std. Dev. 2.38); PCs per employee (non-fin) mean 0.497 (Std. Dev. 0.092).

*Source: wpiea2021214-print-pdf (excerpts provided).*

### 0.  Firms are heterogeneous at date 0 along two publicly observable dimensions.  First,

### wpiea2021214-print-pdf - 0.  Firms are heterogeneous at date 0 along two publicly observable dimensions.  First,

### Model setup: agents, heterogeneity, and project technology
- Firms: unit mass normalized to one; share of young firms is y∈(0,1).  Firm age and collateral are independent.
- Collateral: drawn from continuous distribution G(C).  Market price of collateral at date 1 is P.
- Project payoff: yields x > 1 if successful and 0 if unsuccessful.
- Project quality (private to firm): good (G) with success probability p_G and bad (B) with success probability p_B, where 0 < p_B < p_G < 1 and only good projects have positive NPV: p_B x < 1 < p_G x.
- Share of good projects at date 0 is q > 0, independent of bank or firm characteristics.
- Adverse selection assumption (severe): [q p_G + (1−q) p_B] x < 1, so banks screen all borrowers in equilibrium and as a result all loans granted are to good firms.

### Banks, IT adoption, and screening tools
- Banks: unit mass of banks, each endowed with one unit of investment good at date 0.
- IT adoption: exogenous fraction h∈(0,1) of banks are high-IT; remainder are low-IT.
- Screening tools:
  - Information acquisition: bank pays fixed cost F to learn project type. Fixed cost is lower for old firms than young firms: F_O < F_Y.
  - Collateral-based screening: bank asks for collateral at date 0, repossessed and sold at date 1 if default. Cost of screening via collateral is lower for high-IT banks than low-IT banks: v_HighIT < v_LowIT.
- Lending volume: maximizes joint surplus; banks receive fraction θ∈(0,1) of surplus. Banks and firms are randomly matched; startups do not apply to multiple banks (no multi-application).

### Assumed ranking of screening costs (baseline)
- v_HighIT < F_O < p_G x − 1 < min{F_Y, v_LowIT}.

### Incentive compatibility and minimum collateral
- Only good firms may receive credit in equilibrium (fraction q).
- Young firms receive credit only when matched with a high-IT bank (fraction h of banks) and if collateral C > C_min, where C_min solves:
  - p_B (x − r) ≡ (1 − p_B) P C_min,
  - r is the bank’s lending rate.
- Interpretation: left-hand side is benefit to a bad firm from pretending to be good; right-hand side is cost from losing collateral value upon failure. C_min depends on P: C_min = C_min(P).

### Equilibrium screening for old firms
- Old firms always receive credit.
  - Matched to high-IT bank: lending is backed by collateral if C ≥ C_min, otherwise bank acquires information (pays F_O).
  - Matched to low-IT bank: screening via information acquisition is used.

### Proposition and main comparative-statics (share of lending to young firms)
- Proposition 1: share of lending to young firms is
  - s_Y ≡ y h [1 − G(C_min)] / [1 − y + y h [1 − G(C_min)]].
- Predictions (comparative statics):
  - Prediction 1: ds_Y / dh > 0. A higher share of high-IT banks increases the share of lending to young firms.
  - Prediction 2: ds_Y / dP > 0. A higher collateral value increases the share of lending to young firms.
  - Prediction 3: d^2 s_Y / (dh dP) > 0. A higher collateral value increases the share of lending to young firms more when the share of high-IT banks is higher.
  - Prediction 4: Bank IT adoption does not affect the quality (default rate) of firms receiving funding in equilibrium (all funded projects are good).
- Additional model implication: young firms use more collateral than old firms in equilibrium.

### Extensions and additional predictions
- Extension 1: Recourse versus non-recourse states
  - Fraction i∈(0,1) of firm owners generate external income I ≥ P C_min; banks in recourse states can access I upon default; in non-recourse states they cannot.
  - Collateral and recourse are substitutes for deterring bad firms.
  - Prediction 5: ds_Y / dh is larger in non-recourse than in recourse states:
    - ds_Y / dh | Non−recourse > ds_Y / dh | Recourse.
  - Similarly, the interaction effect with collateral is stronger in non-recourse states:
    - d^2 s_Y / (dh dP) | Non−recourse > d^2 s_Y / (dh dP) | Recourse.
- Extension 2: Distance (soft information and proximity)
  - Allow low-IT banks to screen some young firms that are close by by assuming:
    - F_close^Y < p_G x − 1 < F_distant^Y < v_LowIT.
  - Prediction 6: Distance matters less for high-IT banks than for low-IT banks; high-IT bank lending is relatively insensitive to distance shocks, low-IT bank lending is more sensitive and favors close young firms.

### Empirical implementation: data construction and variables
- Sample period: 1999–2007 (main analysis), chosen to avoid post-crisis regulatory changes.
- Bank IT adoption measure:
  - Source: establishment-level survey on personal computers per employee by CiTBDs Aberdeen for years 1999, 2003, 2004, and 2006; focus on banking sector establishments (SIC2), 143,607 establishment-year observations.
  - Construct bank-level IT measure ̃IT_b by regressing PCs/Emp on a bank fixed effect and controls (establishment type, county, year, log employees) for years 1999, 2003, 2004, 2006.
  - County-level IT exposure IT_c constructed using branch distribution in 1994:
    - IT_c = Σ_{b=1}^N ̃IT_b * No.Branches_{b,c} / No.Branches_c,
    - standardized to mean zero and standard deviation one.
  - Correlations: IT budget (2016) correlates with PCs per employee at 0.65; cross-sectional R^2 of PCs per Employee on per capita IT budget is 44%.
- Young firm measure:
  - Source: Quarterly Workforce Indicators (QWI), employment at county-two-digit NAICS industry-year level by firm age brackets.
  - Define young firms as age 0–1. Use 4th quarter values; employment of young firms is a flow (jobs created by new firms).
  - Baseline normalization: job creation of young firms scaled by total employment in same county-industry cell.
- Industry-level home equity dependence:
  - Source: 2007 Public Use Survey of Business Owners (SBO); compute fraction of young firms reporting use of home equity or personal assets to start/expand, aggregated by 2-digit industry.
- County controls: log total population, share black population, share population older than 65, unemployment rate, house price growth (FHFA HPI), log per capita income.
- Bank balance-sheet data: FDIC SDI (second quarter each year) for assets, Tier 1, non-interest & total income, investment securities, efficiency ratio, non-performing loans, ROA, deposits.
- Small business lending response: CRA data on loan originations below $1 million by financial institutions with > $1 billion assets, aggregated to BHC-county level. Year-to-year change normalized by midpoint:
  - ∆CRA_{b,c,t} = (CRA_{b,c,t} − CRA_{b,c,t−1}) / ((CRA_{b,c,t} + CRA_{b,c,t−1})/2),
  - bounds growth in [−2, 2].

### Main empirical tests and findings (tests of model predictions)
- Empirical specification for Prediction 1 (cross-sectional at county-industry level):
  - startups_{c,i} = β_1 IT_exposure_{c,99} + β_2 constraint_i + β_3 IT_exposure_{c,99} × constraint_i + controls_{c,99} + θ_c + φ_i + ε_{c,i},
  - startups_{c,i} is employment share of firms age 0–1 out of total employment, averaged 1999–2007.
- Reduced-form evidence:
  - Nonparametric plot (Figure 2) shows positive relation between IT exposure and startup employment.
  - Table 3 regressions:
    - Column (1): positive relation between county IT exposure and startup share.
    - Column (2): coefficient stable after county-level controls.
    - Column (3): including NAICS2 industry fixed effects leaves coefficient statistically and economically similar; one standard deviation higher IT exposure associated with 0.38 pp increase in share of young firm employment (4% of mean 9.3%).
    - Column (4): interaction with industry external-finance dependence positive and significant; a one standard deviation higher IT exposure associated with 1 pp increase in young firm employment share in industries dependent on external finance (11% of mean).
    - Column (5): adding county fixed effects yields similar results (interaction estimate changes by 0.02 pp).
  - Results interpreted as consistent with Prediction 1 and a lending channel: IT exposure matters more in industries that depend on external finance.
- Robustness exercises (Table A1 highlights):
  - Alternative IT exposure constructions (unweighted average, deposits-weighted) produce consistent results.
  - Excluding startups in financial and education industries; excluding Wyoming (high IT exposure); including state fixed effects; alternative normalization; results robust.
  - Stronger relationship within tradable industries (manufacturing and mining/extraction aggregated to 2-digit) — suggests not driven by local demand and consistent with financing channel.
  - Excluding top VC counties or seven high-VC states yields similar results — not driven by VC concentration.
  - Coverage concerns: average survey coverage is 13.6% (sd 8.4%); dropping bottom coverage quartile or including coverage as control does not change results.
- Instrumental variables approach:
  - Two-stage IV using predicted bank deposits from gravity model adjusted by staggered interstate banking deregulation to form ̂IT_c:
    - ̂IT_c = Σ_{b=1}^N ̃IT_b * ̂Deposits_{b,c} / Deposits_c.
  - IV results (Table 4): first stage shows positive correlation between actual and predicted IT exposure; reduced form positive; second stage IV impact qualitatively similar and larger in magnitude than OLS. Difference between OLS and IV estimates not statistically different from zero.
- Changes over time (difference approach):
  - Define ∆IT_c = Σ_{b=1}^N ∆̃IT_b * No.Branches_{b,c} / No.Branches_c, where ∆̃IT_b is increase in bank IT adoption between 1999 and 2006.
  - Counties more exposed to increases in IT experienced less negative decreases in startup rates (Figure 3); formal regressions in Table A2 confirm positive correlation between ∆IT and changes in startup rates.

*Italic: Source: wpiea2021214-print-pdf (excerpts provided).*

### 4.2    IT, house prices, and entrepreneurship (Predictions 2 & 3)

### 4.2    IT, house prices, and entrepreneurship (Predictions 2 & 3)

### Theoretical predictions
- Prediction 2: d sY / dP > 0 : a higher collateral value increases the share of lending to young firms.
- Prediction 3: d2 sY / dh dP > 0 : a higher collateral value increases the share of lending to young firms more when the share of high-IT banks is higher.
- Collateral channel emphasized: rising real estate prices increase collateral values, mitigating informational frictions and relaxing borrowing constraints for small and young firms.

### Empirical strategy
- Estimated panel regression at county-industry-year level (1999–2007):
  startups_c,i,t = γ1 IT exposure_c,99 + γ2 ∆HPI_c,t + γ3 IT exposure_c,99 × ∆HPI_c,t + controls_c,t−1 + θ_c,i + τ_t + ε_c,i,t. (Equation 13)
- Dependent variable: employment share of firms age 0–1 out of total employment in county (c) and 2-digit industry (i) in year (t).
- IT exposure_c,99: counties’ IT exposure as of 1999, standardized to mean zero and a standard deviation of one.
- ∆HPI_c,t: yearly county-level growth in house prices.
- Controls (lagged one period unless county fixed effects included): log total population, share of population age 65 and older, share of black population, education, unemployment rate, employment shares in industries 23, 31, 44, 62, and 72, and PCs per employee in non-financial firms.
- Standard errors clustered at the county level.
- Identification: county-industry fixed effects and industry×year fixed effects used to isolate within county-industry variation and absorb the level effect of IT exposure.

### Main empirical findings (tests of Predictions 2 & 3)
- Support for Prediction 2:
  - γ2 > 0 is implied by Prediction 1; empirical estimates show a rise in house prices is associated with an increase in local entrepreneurship conditional on year fixed effects.
  - Finding persists when controlling for county IT adoption.
- Support for Prediction 3:
  - Focus on γ3 (interaction of IT exposure and ∆HPI). Estimates find γ3 > 0, consistent with Prediction 3.
  - Interaction coefficient remains positive and similar in size when adding time-varying county controls and industry×year fixed effects.
- Robustness:
  - County-industry fixed effects absorb IT exposure level; interaction captures how house-price sensitivity of entrepreneurship varies with initial IT exposure.

### Industry heterogeneity
- Prior literature: young firms are more responsive to collateral value changes in industries with lower average start-up capital or greater reliance on home equity.
- Empirical approach: exploit industry heterogeneity within county and year (controls include industry×year and county×year fixed effects).
- Results: Larger benefits of house price increases due to presence of high-IT banks occur predominantly in industries whose financing is more sensitive to collateral-value changes (columns (6) and (7) of Table 5).

### Complementary evidence and extensions
- Appendix A1 (bank-county-year level): high-IT banks’ small business lending reacts more to increases in local house prices.
- Summary interpretation: entrepreneurship increases when local collateral values increase, especially in counties with higher exposure to IT-intensive banks — consistent with Predictions 1–3.

### Implications highlighted in this section
- IT in banking amplifies the collateral channel: banks with higher IT adoption are more likely to translate increases in home values into lending to young firms.
- Policy relevance:
  - IT adoption in lending decisions can spur job creation by young firms by making collateralized lending cheaper.
  - With the strong rise in house prices since the pandemic and greater reliance on IT due to reduced physical interactions, IT adoption in banking can spur entrepreneurship and productivity growth in the post-pandemic world.

*Source: wpiea2021214-print-pdf - 4.2    IT, house prices, and entrepreneurship (Predictions 2 & 3)*

### References

### wpiea2021214-print-pdf - References

### Main empirical findings
- County-level IT exposure of banks is positively associated with entrepreneurship (share of employment in firms age 0-1).
  - Table 3: IT exposure coefficients: 0.455***, 0.397***, 0.370***, 0.373*** (standard errors in parentheses: (0.118), (0.098), (0.098), (0.098)); Observations: 25,742; R-squared: 0.003, 0.047, 0.252, 0.252, 0.354 across columns.
- Instrumental variables evidence supports the positive relationship between county IT exposure and startup employment.
  - Table 4: OLS and IV results: IT exposure (OLS) 0.319*** (0.109); IT exposure (second stage) 0.526*** (0.143); Gravity/RS first-stage coefficients 0.640*** (0.0667) and 0.337*** (0.0889); Observations: 19,293; R-squared reported: 0.246, 0.536, 0.247, 0.051.
- Interaction with housing collateral and home equity:
  - Table 5: ∆ HPI coefficients include 0.025**, 0.024**, -0.024**, -0.041***, -0.034***, -0.028** (standard errors reported); IT exposure×∆ HPI coefficients include 0.075***, 0.070**, 0.075**, 0.271*** (standard errors: (0.027), (0.033), (0.030), (0.086)).
  - Additional interactions: IT exposure×∆ HPI×Low SU capital 0.136*** (0.051); IT exposure×∆ HPI×home equity 0.175** (0.087); IT exposure×∆ HPI×Recourse -0.264*** (0.092).
- Transition rates of young firms show no strong positive effect of IT exposure on survival transitions.
  - Table 6: IT exposure coefficients approximately 0.000 (standard errors 0.000); IT exposure×ext. fin. dep coefficients around -0.001 (standard errors 0.001); Observations: 23,696 (columns 1–3) and 22,643 (columns 4–6); R-squared: 0.070, 0.070, 0.140, 0.048, 0.048, 0.120.
- Recourse law heterogeneity:
  - Table 7: IT exposure coefficients across specifications: 0.305*** (0.0966), 0.471*** (0.176), 0.700*** (0.203), 0.673*** (0.204). Interaction Recourse State×IT exposure coefficients: -0.463** (0.220), -0.434** (0.220). Observations vary: 20,046; 5,696; 25,742; 24,630.
- Banks’ IT and CRA/HE lending responsiveness to house prices:
  - Table A5 (CRA loans): IT×∆ HPI coefficients: 0.213*** (0.066), 0.244*** (0.068), 0.310*** (0.103), 0.178 (0.147) across specifications; IT main effect examples: 0.031*** (0.002), 0.007* (0.004), 0.006 (0.004). Observations up to 338,857; R-squared range: 0.028 to 0.331.
  - Table A6 (Home equity loans): ∆ HPI coefficients: 2.158*** (0.090), 2.811*** (0.231), 5.013*** (0.408), 2.248*** (0.106); IT×∆ HPI coefficients: 0.343*** (0.085), 0.376*** (0.090), 0.465*** (0.143), 0.931*** (0.150); IT main effects negative examples: -0.022*** (0.005), -0.066*** (0.010), -0.061*** (0.010). Observations: up to 50,036; R-squared up to 0.633.

### Key numerical summaries and descriptive statistics
- Table 1 (county-level descriptive statistics; Obs = 1774 unless noted):
  - IT exposure: Mean -0.001, Std. Dev. 1.235, Min -0.562, Max 0.964, P25 -0.108, P50 -0.041, P75 0.067.
  - log(pop): Mean 10.995, Std. Dev. 1.135, Min 8.50, Max 16.06, P25 10.186, P50 10.774, P75 11.651.
  - log(income pc): Mean 10.062, Std. Dev. 0.206, Min 9.493, Max 11.305, P25 9.929, P50 10.039, P75 10.163.
  - bachelor or higher mean 0.183 (Std. Dev. 0.083); share pop old mean 0.138 (Std. Dev. 0.037); share pop black mean 0.091 (Std. Dev. 0.133).
  - unemployment rate mean 4.67 (Std. Dev. 2.38).
  - PCs per employee (non-fin) mean 0.497 (Std. Dev. 0.092).
- Table 2 (balancedness split by low/high IT; Observations 592 / 591 / 1183):
  - log(pop) low IT mean 10.94 (1.11), high IT mean 10.82 (1.10), mean diff t = 2.00.
  - log(income pc) low 10.05 (0.20), high 10.04 (0.21), mean diff t = 1.09.
  - PCs per employee low 0.50 (0.10), high 0.49 (0.09), mean diff t = 1.04.
- Table 8 (CRA lending – distance, bank-county-year regressions; Observations vary up to 194,771):
  - ∆ income coefficient example: 0.019*** (0.003).
  - log(distance) coefficients: 0.016*** (0.003), 0.018*** (0.003), 0.055*** (0.005), -0.003 (0.005), 0.017*** (0.003), 0.017*** (0.003).
  - Interaction ∆ income×log(distance) coefficients include -0.003***, -0.004***, -0.009***, 0.002*, -0.004***, -0.003*** (standard errors 0.001).
  - IT main and interactions examples: IT 0.060*** (0.014); ∆ income×IT -0.014*** (0.003), -0.011*** (0.003); IT×log(distance) -0.009*** (0.003), -0.011*** (0.003); triple interaction ∆ income×log(distance)×IT 0.003*** (0.001), 0.002*** (0.001).

### Robustness and additional evidence
- Robustness table (Table A1) shows baseline and alternative specifications:
  - Baseline IT exposure coefficient 0.377*** (0.098), alternative specifications produce coefficients: 0.163** (0.073), 0.398*** (0.106), 0.375*** (0.099), 0.333*** (0.092), 0.418*** (0.126), 0.054 (0.065), 0.809* (0.421), 0.247*** (0.088), 0.349*** (0.095), 0.344*** (0.097), 0.405*** (0.103).
  - IT exposure (deposit weighted) 0.342*** (0.094).
  - Observations range across columns; standard errors clustered at the county level.
- Syndicated loan-level evidence (Table A3) on secured loans:
  - Bank IT coefficients: 0.230*** (0.051), 0.279*** (0.057), 0.039* (0.022), 0.046** (0.019), 0.033* (0.017) across samples; Observations up to 211,796; R-squared values include 0.018, 0.049, 0.820, 0.824, 0.822.

### Mechanisms and interpretation from Appendix A1
- Regression (Equation 16) at bank-county-year level: ∆loans_{b,c,t} = β1 IT_b + β2 ∆HPI_{c,t} + β3 IT_b × ∆HPI_{c,t} + controls + τ_t + ε_{b,c,t}.
- Finding: high-IT banks’ small business lending is more responsive to local house price increases, consistent with greater reliance on hard information or collateral sensitivity (expectation β3 > 0).
- Column-level results: high-IT banks show about 50% larger association between house price growth and loan growth in split-sample comparisons; interaction estimates robust to county×year, bank×county, and bank×year fixed effects.
- Caveat: CRA data cover small firms (most young firms are small but not all small firms are young).

### Figures and notable graphical summaries
- Figure 2: Job Creation by Young Firms and Banks’ IT adoption: reported binscatter with β = 2.50, t = 7.18 illustrating positive relationship between share of employment by young firms and county IT exposure.
- Figure 3: Binscatter of change in startup rate (2006 vs 2000) against change in county IT adoption (standardized).
- Figure 4: Binscatter of young-firm employment share vs county IT exposure split by industries with above/below median home equity usage; blue triangles denote areas where house prices rose above the median and red dots where house prices rose below the median.
- Appendix figures: distribution of share of loans in county with a branch by bank (Figure A1); share of secured loans in Dealscan vs banks’ IT adoption (Figure A2).

*Italic source attribution: References and accompanying figures/tables and Appendix from wpiea2021214-print-pdf - References*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021214-print-pdf.pdf_