## wpiea2020014-print-pdf

## Source details

**Canonical URL:** [wpiea2020014-print-pdf](https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020014-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2020/english/wpiea2020014-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2020/english/wpiea2020014-print-pdf.pdf.json)

---

### Main findings on IT adoption and bank resilience
- Pre-crisis cross-sectional comparison: "This represents a 10% reduction with respect to the cross-sectional average and 14% of the cross-sectional standard deviation."
- Crisis-period effect:
  - "once the crisis hit, a one standard deviation increase in IT adoption could have lowered by 15% the surge in NPLs with respect to pre-crisis levels."
  - Panel estimates: a one standard deviation higher IT adoption is associated with a between 13 and 17 basis points lower NPL share increase during the crisis.
  - Average share of NPLs in the crisis period: 1.5 percent; standard deviation: 1.13.
    - A one standard deviation higher IT adoption led to a reduction in NPLs between 9 and 11% with respect to the mean.
    - The same corresponds to a reduction between 12 and 15% with respect to the cross-sectional standard deviation.
  - The increase between the pre-crisis average and the crisis NPL share is 1.05 percentage points.
    - A uniform one standard deviation increase in IT adoption across all banks would have diminished the surge in NPLs between 12 and 16% (ignoring heterogeneity, spillovers, and general equilibrium effects).
- No significant correlation between pre-crisis IT adoption and banks’ non-performing loans outside the crisis.
- IT adoption is not significantly correlated with banks’ ex-ante exposure to the GFC in terms of geographical footprint or business model as measured by funding sources, assets composition, employees’ wages, and other balance sheet characteristics.
- Banks led by more IT-savvy executives adopted more IT and had fewer NPLs during the crisis, even after controlling for human capital.

### Data sources, measurement, and key statistics
- IT data:
  - Establishment survey on personal computers per employee by CiTBDs Aberdeen for years 1999, 2003, 2004, 2006, and 2016.
  - Sample after cleaning: 143,607 establishment-year observations.
  - Correlation between per-capita PCs and IT budget in 2016: 65%.
  - Cross-sectional regression R-squared of PCs per Employee on per capita IT budget: 44%.
- Matching: "We map 90% of the assets from the bank-level dataset to the IT data."
- Bank-level IT adoption measure:
  - Regression (1999, 2003, 2004, 2006): PCs/Emp_i,t = IT̃_b + θ_type + θ_c + θ_t + γ·Emp + ε_i,t (PCs/Emp capped at top 1%).
  - Regression R-squared: 42%.
  - Share of variation explained: bank fixed effect 60%; year fixed effect 11%; location 27%; number of employees and bank types nearly zero.
  - IT_b is the standardized bank fixed effect (̃IT_b divided by its standard deviation after subtracting its mean).
  - Robustness: results unaffected by aggressive winsorization (5% on both sides) of IT_b.
- NPLs definition (baseline, following Hirtle et al. (2018)):
  - Components: past due 90 days or more and still accruing (bhck5525), nonaccrual (bhck5526), debt securities past due 90 days or more and still accruing (bhck3506), debt securities nonaccrual (bhck3507).
  - Main dependent variable: amount of NPLs scaled by total assets.
  - Winsorization: bank-level ratios winsorized at top 2.5 percent before taking averages (robust to different treatments of outliers).
  - Crisis distribution: "Most banks have an NPL ratio of around 1% in the crisis period, but there is a long right tail ... For some banks almost 5% of their balance sheet consists of NPLs."
- Freddie Mac Single Family Loan-Level Dataset:
  - Loan-level fields used: postal code, credit score FICO, loan-to-value LTV, debt-to-income DTI, origination year, seller, delinquency status.
  - Delinquency defined as past due more than 90 days.
  - Due to seller disclosure limits, only 22 banks have information on technology adoption in the Freddie Mac merge.

### Empirical strategy and robustness
- Panel specification (Equation 2):
  - NPLb,t = αb + δt + β ITb · crisis t + (Xb · crisis t)′ γ + εb,t
  - Sample: observations for years 2001–2014; final sample: 4,608 observations on 337 banks.
  - Estimation via OLS with standard errors double-clustered at the bank and year level (unless otherwise noted).
  - Bank fixed effects included; ITb and Xb appear interacted with the crisis dummy.
- Controls included: loans-to-assets ratio, log(total assets), share of capital over assets, share of wholesale funding over assets, ROA, log(average wages), exposure to house price drop, IT adoption of other banks in same location.
- Tests for omitted variable bias:
  - Inclusion of observable controls does not affect estimated impact of IT on NPLs.
  - Formal testing following Altonji et al. (2005) and Oster (2019) finds "no evidence of a sizeable bias" from unobservable bank-level characteristics.
- Robustness checks (Table A1 highlights):
  - Alternative IT measure (PCs per Emp): results hold.
  - Winsorization of IT and NPLs: results robust.
  - Alternative NPL definitions (past due 60 days): results robust.
  - Alternative clustering and denominators: results robust.
  - Selected IT-adoption×crisis coefficients and standard errors:
    - Baseline: -0.165** (0.068)
    - PCs per Emp: -0.243* (0.120)
    - HW I T winsorized: -0.158** (0.069)
    - HW NPLs winsorized: -0.161** (0.063)
    - Loans Broad def.: -0.242** (0.095)
    - As of 2006: -0.214** (0.080)
    - Bank Clustering: -0.380* (0.183)
    - Alternative clustering: -0.165*** (0.051)

### Time-varying and cross-sectional evidence
- Year-by-year (Equation 3) summary:
  - Effect insignificant in pre-crisis period (1996–2007) except small negative effect significant at the 5% level in 2002.
  - Between 2007 and 2010 the effect is negative and statistically significant at the 5% level; in 2009 and 2010 significant at the 1% level.
  - Coefficient reaches maximum in 2010 at -0.3:
    - Interpretation: a one standard deviation higher IT adoption was associated with 30 basis points lower NPLs in 2010.
  - Impact remains negative in 2011 and 2012 but not statistically significant; no detected impact in 2013 and 2014.
- Cross-sectional regressions (Table 2):
  - IT-adoption coefficient on NPLs (selected): -0.183*** (0.061) in one specification.
  - Sample means (first column): Mean 1.546; Std.Dev. 1.131.

### Channels and mechanisms
- Screening at origination:
  - Mortgages originated before 2007 and sold to Freddie Mac by high-IT adoption banks were significantly less likely to be delinquent during the GFC, implying better borrower screening at origination.
- No evidence of risk-shifting via securitization:
  - Analysis finds no evidence that IT adopters offloaded low-quality loans to GSEs.
- No detectable local spillovers:
  - Adding IT of local competitors as a control: banks in areas where competitors adopted more IT did not suffer a stronger increase in NPLs.
- Role of hard information:
  - Inclusion of credit-score, DTI, LTV does not eliminate IT adoption effect, suggesting either additional information used by IT-adopters or more sophisticated use of available variables.

### Executives’ background and IT adoption
- Hypothesis: Top executives with a more tech-prone background promote higher IT adoption and better outcomes.
- Measurement:
  - Executives considered: CEOs, CFOs, COOs, and Presidents hired before 2007.
  - Tech-orientation measure: search for tech-related keywords in biographies; occurrences scaled by biography length and averaged across top executives to form bank-level executive IT-intensity.
  - Matched sample size: 249 banks.
- Main findings (Table 4 and robustness):
  - Executives’ "tech-orientation": -0.155*** (0.047) associated with NPLs in one reported column; IT-adoption first-stage: 0.0900* (0.051).
  - Column (1) on 249-bank sub-sample finds similar results to main sample (slightly smaller and noisier).
  - Banks led by more tech-oriented executives experienced lower NPLs during the crisis and adopted more IT pre-crisis.
- Instrumental-variable approach:
  - ExecIT used as instrument for IT adoption; first-stage F-stat: 4 (low).
  - IV estimates plausible but weak; IV coefficient larger in magnitude than OLS, likely reflecting weak-instrument finite-sample bias.
- Robustness to compensation and keyword choices:
  - Inclusion of log compensation and non-base compensation share does not affect estimates.
  - Leave-one-word-out tests on keyword list produce coefficients clustered around primary estimates with same sign.

### Bank lending and credit provision
- Descriptive pattern (Figure 3):
  - Share of total loans (normalized by pre-GFC assets) for high- and low-IT adoption groups indistinguishable up to 2006; from 2007 low-IT adopters provided remarkably lower amounts of loans than high-IT counterparts; convergence begins from 2012 but difference remains in 2014.
- Lending regressions (Table 6):
  - A 100 basis points higher NPLs to assets ratio is associated with about 100 basis lower average loan growth.
  - IT-adoption coefficients on loan growth:
    - 0.378** (0.182) in column (1)
    - 0.331* (0.196) in column (3)
  - One standard deviation increase in IT adoption is associated with a 33 basis points higher loan growth during the crisis, about 20% of its mean.
- Conclusion: IT adoption helped banks provide credit during and after the GFC; this may reflect mitigation of delinquency/NPLs and improved operational functioning.

### Limitations, interpretation, and residual concerns
- Technology evolution caveat: "the type of technologies employed by commercial banks in the early 2000s might be different than today’s use of machine learning and big data."
- Data limitations:
  - No direct measures of managerial practices or employees’ education; proxies used (average wages, managers’ compensation).
  - Alternative IT measures (IT budget, cloud computing) present only in later waves; correlation with PCs per employee is high but those waves are not available pre-2007.
- Residual concerns:
  - Potential unobserved organizational or managerial quality correlated with both IT adoption and crisis resilience cannot be completely ruled out, although multiple empirical checks mitigate this concern.
  - Possibility that more "tech-savvy" people are inherently better at creating or sustaining lending practices for other reasons remains.
- Strength: analysis covers a period of severe and systemic turmoil (GFC) and a representative sample covering the vast majority of lending in the pre-crisis period.
- Trade-off: FinTech studies cover more recent technologies but cannot yet observe performance through a comparable systemic shock.

### Policy implications
- Technology adoption in lending can enhance financial stability through better monitoring and screening, reducing the surge in NPLs during system-wide shocks.
- IT adoption did not appear to exacerbate moral hazard via securitization in this context; instead, it had positive aggregate effects without evidence of risk transfer across parties.
- Relevance for ongoing policy debate:
  - Results suggest benefits of IT adoption for resilience in systemic crises, complementary to but not substitutive of the evolving FinTech literature.
  - Contemporary machine learning techniques are framed as more powerful versions of pre-GFC statistical tools; data collection/use (e.g., digital footprint) is conceptually similar to credit scores, with main differences in infrastructure and know-how.

*Source: wpiea2020014-print-pdf (content excerpt).*

### 2010.  This represents a 10% reduction with respect to the cross-sectional average and 14% of the cross-

### wpiea2020014-print-pdf - 2010

### Main findings on IT adoption and bank resilience
- Pre-crisis cross-sectional comparisons: "This represents a 10% reduction with respect to the cross-sectional average and 14% of the cross-sectional standard deviation."
- Crisis-period effect: "once the crisis hit, a one standard deviation increase in IT adoption could have lowered by 15% the surge in NPLs with respect to pre-crisis levels."
- No significant correlation between pre-crisis IT adoption and banks’ non-performing loans outside the crisis.
- IT adoption is not significantly correlated with banks’ ex-ante exposure to the GFC in terms of geographical footprint or business model as measured by funding sources, assets composition, employees’ wages, and other balance sheet characteristics.
- Banks led by more IT-savvy executives adopted more IT and had fewer NPLs during the crisis, even after controlling for human capital.

### Data sources and measurement
- IT data: establishment survey on personal computers per employee by CiTBDs Aberdeen for years 1999, 2003, 2004, 2006, and 2016.
  - Sample after cleaning: 143,607 establishment-year observations.
  - Correlation between per-capita PCs and IT budget in 2016: 65%.
  - Cross-sectional regression R-squared of PCs per Employee on per capita IT budget: 44%.
- Matching: mapped bank branches from Aberdeen dataset to BHC data; "We map 90% of the assets from the bank-level dataset to the IT data."
- Bank-level IT adoption measure:
  - Regression (1999, 2003, 2004, 2006): PCs/Emp_i,t = IT̃_b + θ_type + θ_c + θ_t + γ·Emp + ε_i,t (PCs/Emp capped at top 1%).
  - Regression R-squared: 42%.
  - Share of variation explained: bank fixed effect 60%; year fixed effect 11%; location 27%; number of employees and bank types nearly zero.
  - IT_b is the standardized bank fixed effect (̃IT_b divided by its standard deviation after subtracting its mean).
  - Robustness: results unaffected by aggressive winsorization (5% on both sides) of IT_b.
- NPLs definition (baseline, following Hirtle et al. (2018)):
  - Total loans, leasing financing receivables and debt securities and other assets - past due 90 days or more and still accruing (bhck5525)
  + Total loans, leasing financing receivables and debt securities and other assets - nonaccrual (bhck5526)
  - Debt securities and other assets - past due 90 days or more and still accruing (bhck3506)
  - Debt securities and other assets - nonaccrual (bhck3507)
  - Main dependent variable: amount of NPLs scaled by total assets.
  - Winsorization: bank-level ratios winsorized at top 2.5 percent before taking averages (robust to different treatments of outliers).
  - Crisis distribution: "Most banks have an NPL ratio of around 1% in the crisis period, but there is a long right tail ... For some banks almost 5% of their balance sheet consists of NPLs."
- Freddie Mac Single Family Loan-Level Dataset:
  - Loan-level fields used: postal code, credit score FICO, loan-to-value LTV, debt-to-income DTI, origination year, seller, delinquency status.
  - Delinquency defined as past due more than 90 days.
  - Due to seller disclosure limits, only 22 banks have information on technology adoption in the Freddie Mac merge.

### Empirical strategy and robustness checks
- Bank fixed effects estimate of IT adoption controls for geography (county fixed effects), establishment type, and size (log employees).
- Tests for omitted variable bias:
  - Inclusion of observable controls (funding sources, assets composition, wages, balance sheet characteristics) does not affect the estimated impact of IT on NPLs.
  - Formal testing following Altonji et al. (2005) and Oster (2019) finds "no evidence of a sizeable bias" from unobservable bank-level characteristics.
- Controls in loan-level analysis:
  - Postal code and origination-year fixed effects included.
  - Controls for borrower credit-score, debt-to-income ratio, and loan-to-value ratio: results on delinquency remain unchanged in linear regression and probit models.
- Executive text-analysis:
  - Biographies searched for tech-related keywords (technology, engineering, math, computer, machine, system, analytic, technique, method, process, stem, efficiency, efficient, software, hardware, data, informatic).
  - Occurrences scaled by biography length and averaged across top executives to form bank-level executive IT-intensity.
  - Finding: banks with higher executive tech-orientation adopted more IT and had lower NPLs during the crisis; effect stable to controls for managers’ compensation and non-base compensation share.

### Channels and mechanisms
- Screening at origination:
  - Mortgages originated before 2007 and sold to Freddie Mac by high-IT adoption banks were significantly less likely to be delinquent during the GFC, implying better borrower screening at origination.
- No evidence of risk-shifting via securitization:
  - If IT adopters had simply offloaded bad loans to GSEs, IT intensity would not enhance financial stability; the analysis finds no evidence that IT adopters offloaded low-quality loans to GSEs.
- No detectable local spillovers where high-risk individuals substituted toward low-IT banks.
- Role of hard information:
  - Inclusion of credit-score, DTI, LTV does not eliminate IT adoption effect, suggesting either (a) additional information used by IT-adopters or (b) more sophisticated use of available variables.

### Additional empirical patterns
- Banks with less pre-crisis IT adoption and banks with higher NPLs during the crisis exhibited significantly weaker loan growth in the crisis.
- Other correlates of higher NPLs during the crisis:
  - Larger banks, banks with more loans, banks with more wholesale funding, and banks with greater geographical exposure to house price shocks had more NPLs.
- House price exposure measure:
  - County-level home value index from Zillow.
  - Constructed percentage change in annual average house price decrease between peak (2012 Q3) and trough (2007 Q4).
  - Bank-level exposure: median decrease across establishments for each bank.

### Limitations and interpretation
- Technology evolution caveat: "the type of technologies employed by commercial banks in the early 2000s might be different than today’s use of machine learning and big data."
- Trade-offs highlighted:
  - Strength: analysis covers a period of severe and systemic turmoil (GFC) and a representative sample covering the vast majority of lending in the pre-crisis period.
  - Limitation: FinTech studies cover more recent technologies but cannot yet observe performance through a comparable systemic shock.
- Data limitations:
  - No direct measures of managerial practices or employees’ education; proxies used (average wages, managers’ compensation).
  - Alternative IT measures (IT budget, cloud computing) present only in later waves; correlation with PCs per employee is high but those waves are not available pre-2007.
- Residual concerns:
  - Potential unobserved organizational or managerial quality correlated with both IT adoption and crisis resilience cannot be completely ruled out, although multiple empirical checks mitigate this concern.
  - Possibility that more "tech-savvy" people are inherently better at creating or sustaining lending practices for other reasons remains.

### Policy implications
- Technology adoption in lending can enhance financial stability through better monitoring and screening, reducing the surge in NPLs during system-wide shocks.
- IT adoption did not appear to exacerbate moral hazard via securitization in this context; instead, it had positive aggregate effects without evidence of risk transfer across parties.
- Relevance for ongoing policy debate:
  - Results suggest benefits of IT adoption for resilience in systemic crises, complementary to but not substitutive of the evolving FinTech literature.

*Source: wpiea2020014-print-pdf - 2010 (content excerpt).*

### 4.1  Panel

### 4.1 Panel

### Empirical specification and data
- Panel equation estimated (Equation 2):
  - NPLb,t = αb + δt + β ITb · crisis t + (Xb · crisis t)′ γ + εb,t
  - NPLb,t: share of non-performing loans relative to assets for bank (BHC) b in year t.
  - ITb: bank-level measure of IT adoption before the crisis (as defined in section 3).
  - αb and δt: bank and year fixed effects.
  - Xb: vector of pre-crisis bank characteristics (simple averages between 2001 and 2006): ratio of loans to assets, log(total assets), share of capital over assets, share of wholesale funding over assets, ROA, log(average wages).
  - Two geographic controls: exposure to the house price drop (county-level drop weighted by number of branches) and IT adoption of other banks operating in the same location.
- Sample and estimation:
  - Observations for years 2001–2014, keeping only observations with all Xb variables.
  - Final sample: 4,608 observations on 337 banks.
  - Since bank fixed effects are included, ITb and Xb appear only interacted with the crisis dummy.
  - Estimation via OLS with standard errors double-clustered at the bank and year level (unless otherwise noted).

### Main empirical findings
- Baseline and interaction effects:
  - The base effect of IT adoption on NPLs in normal times: negative but not statistically significant.
  - Interaction (IT adoption × crisis dummy): negative and statistically significant.
  - Interpretation: during the crisis (defined as 2007–2010), banks that adopted more IT before the crisis had a significantly lower share of NPLs than banks with less IT adoption.
- Magnitude of the crisis effect:
  - A one standard deviation higher IT adoption is associated with a between 13 and 17 basis points lower NPL share increase during the crisis.
  - The average share of NPLs was 1.5 percent in the crisis period; standard deviation was 1.13.
    - A one standard deviation higher IT adoption led to a reduction in NPLs between 9 and 11% with respect to the mean.
    - The same corresponds to a reduction between 12 and 15% with respect to the cross-sectional standard deviation.
  - The increase between the pre-crisis average and the crisis NPL share is 1.05 percentage points.
    - A uniform one standard deviation increase in IT adoption across all banks would have diminished the surge in NPLs between 12 and 16% (ignoring heterogeneity, spillovers, and general equilibrium effects).

### Controls and heterogeneity
- Controls added progressively (Columns (5)–(12)):
  - Pre-crisis loans-to-assets ratio:
    - Banks with more loans as a share of assets had a stronger increase in NPLs.
  - Exposure to house price drops (branch-weighted county peak-to-trough):
    - Banks with more branches in counties with larger house price drops suffered a stronger increase in NPLs.
  - Bank size (log assets) interacted with crisis:
    - Larger banks had a stronger increase in NPLs in the crisis (consistent with Sullivan and Vickery (2013)).
  - Pre-crisis capital position, pre-crisis wholesale funding, ROA, and average wage:
    - These did not have a significant impact on NPLs in the crisis in the panel regressions.
- Local-competitor IT adoption (spillover test):
  - Adding IT of local competitors as a control tests for negative spillovers (borrower rejected by high-IT bank applying to low-IT bank).
  - Column (12): banks based in areas where competitors adopted more IT did not suffer a stronger increase in NPLs relative to banks where local competitors did not adopt IT intensively.
  - Evidence suggests IT adoption does not have negative spillover effects to local competitors.

### Robustness checks
- Table A1 (summary of robustness tests):
  - Column (1): repeats baseline.
  - Column (2): alternative IT measure — average share of PCs per employee across branches for each bank; results hold.
  - Column (3) and (4): stronger winsorization of the IT adoption measure and NPLs; results robust to outliers.
  - Column (5) and (7): alternative denominators for dependent variable:
    - Column (5): divide by overall loans of the bank.
    - Column (7): divide by pre-crisis assets.
  - Column (6): alternative NPL definition — past due 60 days or more (broader classification vs. baseline 90 days); results robust.
  - Column (8): standard errors clustered at the bank-level instead of double clustered at bank and year; results reported.

### Time-varying effects (year-by-year analysis)
- Year-varying specification (Equation 3):
  - NPLb,t = αb + δt + Στ≠2006 βτ · ITb · 1[t=τ] + εb,t, with coefficient for 2006 normalized to zero.
- Summary of estimated βτ (Figure 2 description):
  - Effect of IT adoption on NPLs is insignificant in pre-crisis period (1996–2007) except a small negative effect statistically significant at the 5% level in 2002.
  - Between 2007 and 2010 the effect is negative and statistically significant at the 5% level; in 2009 and 2010 the effect is statistically significant at the 1% level.
  - Coefficient reaches maximum in 2010 at -0.3:
    - Interpretation: a one standard deviation higher IT adoption was associated with 30 basis points lower NPLs in 2010.
  - Impact remains negative in 2011 and 2012 but is not statistically significant.
  - No detected impact in 2013 and 2014.

### Interpretation and implications
- Evidence indicates that pre-crisis IT adoption materially reduced NPL increases during the crisis years.
- The stability of the coefficient across specifications suggests low correlation between included controls and the IT adoption measure.
- Lack of negative spillovers to local competitors implies IT adoption reduced NPLs for adopters without simply shifting bad loans to non-adopters in the same geographic area.
- Quantitatively meaningful effect: one standard deviation in IT adoption leads to a sizable reduction in bank-level crisis NPL increases (13–17 basis points) and, under a uniform increase scenario, could reduce the aggregate surge in NPLs by 12–16% absent general equilibrium and spillover adjustments.

*Source: 4.1 Panel, wpiea2020014-print-pdf*

### 5.2  Executives’ Background

### 5.2  Executives’ Background

### Motivation and hypothesis
- Question: Why do some banks adopt less IT than others despite its beneficial effects?
- Conjecture: Top executives with a more tech-prone background and orientation may overcome frictions (information, incentives, financial constraints) and promote higher IT adoption in the banks they lead.
- Empirical motivation: Explained variation in technology adoption at the branch-level is driven by bank characteristics (60%) relative to geographic characteristics (27%).

### Data and measurement
- Executives considered: CEOs, CFOs, COOs, and Presidents hired before 2007 (independently of whether they are still active).
- Tech-orientation measure: Search for tech-related keywords in executives’ biographies; compute an overall "tech-intensity" score for each bank.
- Matched sample size: 249 banks (regulatory data, IT adoption, and executive biographies matched).

### Empirical specification
- Cross-sectional regression estimated:
  - Y_b = α + β · ExecIT_b + ε_b  (Equation (6))
  - Where ExecIT_b is the standardized tech-orientation of bank b’s executives; dependent variable Y_b is either pre-crisis IT adoption or NPLs over assets during the crisis. Both independent variables are standardized to have mean zero and variance one.

### Main findings (Table 4 summaries)
- Column (1): Baseline specification on the 249-bank sub-sample finds similar results to the main sample (slightly smaller and noisier).
- Column (2): Banks led by more tech-oriented executives experienced lower NPLs during the crisis.
- Column (3): Executives’ background is a significant predictor of bank IT adoption.
- Interpretation: Columns (2) and (3) are consistent with the hypothesis that tech-prone executives led banks to adopt IT more intensively and to experience related benefits (lower NPLs) afterward.
- Measurement note: The lower statistical significance in Column (3) vs (2) may reflect different data sources: Column (2)’s LHS variable is measured from regulatory data rather than a survey and likely contains less measurement error.

### Instrumental-variable approach and limitations
- Under the assumption that executives’ tech-orientation affects banks only through IT adoption, ExecIT is used as an instrument for IT adoption in the main specification (Column (1)).
- First-stage performance: Regression of executive tech-orientation on IT adoption delivers a low R-square and an F-stat of 4.
- Implication: The instrument is plausible but weak; IV estimates likely suffer from sizeable finite-sample bias due to the “weak instrument” problem.
- Reported IV result (Column (4)): The estimated coefficient of IT adoption is negative and statistically different from zero but of an order of magnitude larger than OLS, possibly reflecting poor small-sample properties.

### Robustness checks addressing alternative explanations
- Human capital / compensation concern:
  - Re-estimate Equation (6) including (log of) pre-crisis compensation of executives as a control (Table 5).
  - Result: Estimates are absolutely unaffected by inclusion of compensation.
  - Compensation itself has no explanatory power for NPLs during the crisis nor for IT adoption before the crisis.
  - Interpretation: If compensation proxies human capital (Becker, 2009), results indicate that tech-orientation—rather than general quality/skills—matters.
  - Results are also unaffected by inclusion of the non-base share of total compensation (can shape risk-taking incentives), as reported in Table A2.
- Ad-hoc word list concern:
  - For each word in the tech-keyword list, compute an additional tech-orientation measure excluding that word; re-estimate regressions.
  - Findings plotted in Figure A4: coefficients clustered around Table 4 estimates (flagged by a dashed line), fairly close, and all have the same sign (negative in top and bottom panels and positive for the mid panel).
  - Interpretation: Results robust to exclusion of any single word from the list.

### Interpretation and contribution
- The analysis links heterogeneity in IT adoption to executives’ tech-orientation, supporting a causal interpretation that IT adoption leads to lower NPLs during the crisis rather than unobserved managerial quality.
- The results are informative about roots of dispersion in IT adoption and provide strong support for IT as a cause of lower NPLs during the crisis.

---

### 5.3  Bank Lending (summary linking to executives / IT)
- Descriptive pattern (Figure 3):
  - Share of total loans (normalized by pre-GFC assets) for high- and low-IT adoption groups tracked 2001–2014.
  - Series indistinguishable up to 2006; from 2007 low-IT adopters provided remarkably lower amounts of loans than high-IT counterparts.
  - Convergence begins from 2012 but difference remains in 2014.
  - Suggests heterogeneity in IT adoption (perhaps via impact on NPLs) helps explain different lending dynamics during and after the crisis.
- Formal test specification:
  - ∆Loans_GFC_b = α + β · X_b + ε_b  (Equation (7))
  - X_b is either share of NPLs in crisis period or pre-crisis IT adoption; outcome is change in loans over total assets averaged across the crisis period.
- Key lending results (Table 6):
  - A 100 basis points higher NPLs to assets ratio is associated with about 100 basis lower average loan growth (consistent with Peek and Rosengren (2000)).
  - One standard deviation increase in IT adoption is associated with a 33 basis points higher loan growth during the crisis, which is about 20% of its mean.
- Conclusion: IT adoption helped banks provide credit during (and after) the GFC; whether via mitigation of delinquency/NPLs or via improving operational functioning, IT intensity improved financial stability during the GFC.

### 6  Conclusion (selected points relevant to executives/IT)
- Main paper result: High-IT-adopters experienced a significantly smaller increase in NPLs and provided more credit during the crisis; high- and low-IT-adopters were not differentially exposed pre-crisis in terms of geographic footprint and business model.
- Loans originated by high-IT banks experienced lower delinquency rates during the crisis even when securitized and sold to Freddie Mac—evidence that IT adoption helped select better borrowers and produce more resilient loans.
- Roots of heterogeneity partially related to the “tech-orientation” of top executives, captured via simple text-analysis.
- Caveat: Technologies before the GFC may differ from current fintech/BigTech implementations.
- Supporting considerations for relevance:
  - The simple branch-level measure (computers per employee) correlates with later IT-budget: regression against overall IT-budget of an establishment in 2016 delivers an R-square of 44% and a correlation coefficient of 65%.
  - Many IT-driven changes are recent and untested by a large systemic shock; past shock evidence remains informative.
  - Contemporary machine learning techniques are more powerful versions of pre-GFC statistical tools rather than entirely different systems; data collection/use (e.g., digital footprint) is conceptually similar to credit scores, with main differences in infrastructure and know-how.

*Source: wpiea2020014-print-pdf - 5.2  Executives’ Background*

### References

### References

### Bibliographic citations
- Acharya, Viral V, Philipp Schnabl, and Gustavo Suarez, “Securitization without risk transfer,” Journal of Financial Economics, 2013, 107(3), 515–536.
- Adelino, Manuel, Antoinette Schoar, and Felipe Severino, “Loan originations and defaults in the mortgage crisis: The role of the middle class,” The Review of Financial Studies, 2016, 29(7), 1635–1670.
- Akerman, Anders, Ingvil Gaarder, and Magne Mogstad, “The skill complementarity of broadband internet,” The Quarterly Journal of Economics, 2015, 130(4), 1781–1824.
- Altonji, Joseph G, Todd E Elder, and Christopher R Taber, “Selection on observed and unobserved variables:  Assessing the effectiveness of Catholic schools,” Journal of Political Economy, 2005, 113(1), 151–184.
- Atkin, David, Azam Chaudhry, Shamyla Chaudry, Amit K Khandelwal, and Eric Verhoogen, “Organizational barriers to technology adoption:  Evidence from soccer-ball producers in Pakistan,” The Quarterly Journal of Economics, 2017, 132(3), 1101–1164.
- Autor, David H, Frank Levy, and Richard J Murnane, “The skill content of recent technological change: An empirical exploration,” The Quarterly Journal of Economics, 2003, 118(4), 1279–1333.
- Bai, Jennie, Thomas Philippon, and Alexi Savov, “Have financial markets become more informative?,” Journal of Financial Economics, 2016, 122(3), 625–654.
- Bartlett, Robert, Adair Morse, Richard Stanton, and Nancy Wallace, “Consumer-Lending Discrimination in the Era of FinTech,” Unpublished working paper. University of California, Berkeley, 2018.
- Basten, Christoph and Steven Ongena, “The Geography of Mortgage Lending in Times of FinTech,” Swiss Finance Institute Research Paper, 2019.
- Beaudry, Paul, Mark Doms, and Ethan Lewis, “Should the personal computer be considered a technological revolution? Evidence from US metropolitan areas,” Journal of Political Economy, 2010, 118(5), 988–1036.
- Beccalli, Elena, “Does IT investment improve bank performance? Evidence from Europe,” Journal of Banking & Finance, 2007, 31(7), 2205–2230.
- Becker, Gary S, Human capital: A theoretical and empirical analysis, with special reference to education, University of Chicago press, 2009.
- Benmelech, Efraim and Carola Frydman, “Military ceos,” Journal of Financial Economics, 2015, 117(1), 43–59.
- Bennedsen, Morten, Francisco Perez-Gonzalez, and Daniel Wolfenzon, “Do CEOs Matter?,” NYU Working paper no. FIN-06-032, 2006.
- Berg, Tobias, “Playing the devil’s advocate:  The causal effect of risk management on loan quality,” The Review of Financial Studies, 2015, 28(12), 3367–3406.
- Berg, Tobias, Manju Puri, and Jorg Rocholl, “Loan Officer Incentives, Internal Rating Models and Default Rates,” Review of Finance (forthcoming), 2019.
- Berg, Tobias, Valentin Burg, Ana Gombovi ́c, and Manju Puri, “On the rise of FinTechs–Credit scoring using digital footprints,” The Review of Financial Studies (forthcoming), 2019.
- Berger, Allen N, “The economic effects of technological progress:  Evidence from the banking industry,” Journal of Money, Credit and Banking, 2003, pp. 141–176.
- Bernanke, Ben S, The courage to act:  A memoir of a crisis and its aftermath, WW Norton & Company, 2015.
- Berti, Katia, Christian Engelen, Borek Vasicek et al., “A macroeconomic perspective on non-performing loans (NPLs),” Quarterly Report on the Euro Area (QREA), 2017, 16(1), 7–21.
- Bertrand, M and A Schoar, “The Managing with Style: The Effect of Managers on Corporate Policy,” Quarterly Journal of Economics, 2003, 118(4), 1169.
- Bessen, James E and Cesare Righi, “Shocking Technology:  What Happens When Firms Make Large IT Investments?,” Boston Univ. School of Law, Law and Economics Research Paper, 2019, (19-6).
- BIS, “Resolution of non-performing loans–policy options,” BIS FSI Insights on policy implementation, 2017, 3.
- Bloom, Nicholas and Nicola Pierri, “Cloud Computing Is Helping Smaller, Newer Firms Compete,” Harvard Business Review, 2018.
- Bloom, Benn Eifert, Aprajit Mahajan, David McKenzie, and John Roberts, “Does management matter? Evidence from India,” The Quarterly Journal of Economics, 2013, 128(1), 1–51.
- Bloom, Erik Brynjolfsson, Lucia Foster, Ron Jarmin, Megha Patnaik, Itay Saporta-Eksten, and John Van Reenen, “What drives differences in management practices?,” American Economic Review, 2019, 109(5), 1648–83.
- Bloom, Raffaella Sadun, and John Van Reenen, “Americans do IT better: US multinationals and the productivity miracle,” American Economic Review, 2012, 102(1), 167–201.
- Bresnahan, Timothy F, Erik Brynjolfsson, and Lorin M Hitt, “Information technology, workplace organization, and the demand for skilled labor: Firm-level evidence,” The Quarterly Journal of Economics, 2002, 117(1), 339–376.
- Brynjolfsson, Erik and Lorin M Hitt, “Computing productivity: Firm-level evidence,” Review of Economics and Statistics, 2003, 85(4), 793–808.
- Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru, “Fintech, regulatory arbitrage, and the rise of shadow banks,” Journal of Financial Economics, 2018, 130(3), 453–483.
- Caballero, Ricardo J, Takeo Hoshi, and Anil K Kashyap, “Zombie lending and depressed restructuring in Japan,” American Economic Review, 2008, 98(5), 1943–77.
- Carlin, Bruce I, Arna Olafsson, and Michaela Pagel, “FinTech and Consumer Financial Well-Being in the Information Age,” 2019.
- Claessens, Stijn, Jon Frost, Grant Turner, and Feng Zhu, “Fintech credit markets around the world: size, drivers and policy issues,” BIS Quarterly Review September, 2018.
- D’Acunto, Francesco, Alberto Rossi, and Michael Weber, “Crowdsourcing financial information to change spending behavior,” Chicago Booth Research Paper, 2019, (19-09).
- Degryse, Hans and Steven Ongena, “Distance, lending relationships, and competition,” The Journal of Finance, 2005, 60(1), 231–266.
- Demirgüç-Kunt, Asli and Enrica Detragiache, “Does deposit insurance increase banking system stability? An empirical investigation,” Journal of Monetary Economics, 2002, 49(7), 1373–1406.
- Duval, Romain, Gee-Hee Hong, and Yannick Timmer, “Financial Frictions and the Great Productivity Slowdown,” Review of Financial Studies (forthcoming), 2019.
- Farboodi, Maryam, Adrien Matray, and Laura Veldkamp, “Where has all the big data gone?,” working paper, 2018.
- Finkelstein, Amy, Matthew Gentzkow, and Heidi L Williams, “Place-based drivers of mortality: Evidence from migration,” NBER working paper series, 2019.
- Forman, Chris, Avi Goldfarb, and Shane Greenstein, “The Internet and local wages: A puzzle,” American Economic Review, 2012, 102(1), 556–75.
- FSB, “FinTech and market structure in financial services: Market developments and potential financial stability implications,” Basel: Financial Stability Board, 2019.
- Fuster, Andreas, Matthew Plosser, Philipp Schnabl, and James Vickery, “The role of technology in mortgage lending,” The Review of Financial Studies, 2019, 32(5), 1854–1899.
- Fuster, Paul Goldsmith-Pinkham, Tarun Ramadorai, and Ansgar Walther, “Predictably unequal? the effects of machine learning on credit markets,” The Effects of Machine Learning on Credit Markets (November 6, 2018), 2018.
- Goodman, Laurie S and Jun Zhu, “Loss severity on residential mortgages: evidence from Freddie Mac’s newest data,” The Journal of Fixed Income, 2015, 25(2), 48–57.
- Hau, Harald, Yi Huang, Hongzhe Shan, and Zixia Sheng, “FinTech Credit and Entrepreneurial Growth,” Technical Report, Mimeo, University of Geneva 2018.
- Hau, Harald, Yi Huang, Hongzhe Shan, and Zixia Sheng, “How FinTech Enters China’s Credit Market,” in “AEA Papers and Proceedings,” Vol. 109 2019, pp. 60–64.
- Hirtle, Beverly, Anna Kovner, and Matthew C Plosser, “The impact of supervision on bank performance,” FRB of NY Staff Report, 2018.
- Hughes, Joseph P, Julapa Jagtiani, and Choon-Geol Moon, “Consumer lending efficiency: Commercial banks versus a fintech lender,” Departmental Working Papers, Rutgers University, Department of Economics, 2019.
- Jagtiani, Julapa and Catharine Lemieux, “Fintech lending: Financial inclusion, risk pricing, and alternative information,” 2017.
- Keys, Benjamin J, Amit Seru, and Vikrant Vig, “Lender screening and the role of securitization: evidence from prime and subprime mortgage markets,” The Review of Financial Studies, 2012, 25(7), 2071–2108.
- Keys, Benjamin J, Tanmoy Mukherjee, Amit Seru, and Vikrant Vig, “Financial regulation and securitization: Evidence from subprime loans,” Journal of Monetary Economics, 2009, 56(5), 700–720.
- Keys, Benjamin J, Tanmoy Mukherjee, Amit Seru, and Vikrant Vig, “Did securitization lead to lax screening? Evidence from subprime loans,” The Quarterly Journal of Economics, 2010, 125(1), 307–362.
- Koetter, Michael and Felix Noth, “IT use, productivity, and market power in banking,” Journal of Financial Stability, 2013, 9(4), 695–704.
- Liberti, José María and Mitchell A Petersen, “Information: Hard and soft,” Review of Corporate Finance Studies, 2018, 8(1), 1–41.
- Maggio, Marco Di and Vincent W Yao, “FinTech Borrowers: Lax-Screening or Cream-Skimming?,” Working Paper, 2018.
- Manaresi, Francesco and Nicola Pierri, “Credit supply and productivity growth,” IMF Working Paper, 2019.
- McElheran, Kristina and Chris Forman, “Firm Organization in the Digital Age: IT Use and Vertical Transactions in US Manufacturing,” Available at SSRN 3396116, 2019.
- Meiselman, Ben S, Stefan Nagel, and Amiyatosh Purnanandam, “Judging Banks’ Risk by the Profits They Report,” Working Paper, 2018.
- Mian, Atif and Amir Sufi, “The consequences of mortgage credit expansion: Evidence from the US mortgage default crisis,” The Quarterly Journal of Economics, 2009, 124(4), 1449–1496.
- Mian, Atif and Amir Sufi, “House prices, home equity-based borrowing, and the US household leverage crisis,” American Economic Review, 2011, 101(5), 2132–56.
- Mian, Atif and Amir Sufi, “What explains the 2007–2009 drop in employment?,” Econometrica, 2014, 82(6), 2197–2223.
- Nasiripour, Shahien, “Fintech Lenders Tighten Standards as They Become More Like Banks,” Bloomberg, Dec 16, 2019.
- Navaretti, Giorgio Barba, Giacomo Calzolari, José Manuel Mansilla-Fernandez, and Alberto F Pozzolo, “Fintech and Banking. Friends or Foes?,” Friends or Foes, 2018.
- Oster, Emily, “Unobservable selection and coefficient stability: Theory and evidence,” Journal of Business & Economic Statistics, 2019, 37(2), 187–204.
- Peek, Joe and Eric S Rosengren, “Collateral damage: Effects of the Japanese bank crisis on real activity in the United States,” American Economic Review, 2000, 90(1), 30–45.
- Petersen, Mitchell A and Raghuram G Rajan, “The benefits of lending relationships: Evidence from small business data,” The Journal of Finance, 1994, 49(1), 3–37.
- Petersen, Mitchell A and Raghuram G Rajan, “Does distance still matter? The information revolution in small business lending,” The Journal of Finance, 2002, 57(6), 2533–2570.
- Philippon, Thomas, “On Fintech and Financial Inclusion,” NBER Working Paper, 2019.
- Rajan, Raghuram G, “Has finance made the world riskier?,” European Financial Management, 2006, 12(4), 499–533.
- Rajan, Uday, Amit Seru, and Vikrant Vig, “Statistical default models and incentives,” American Economic Review, 2010, 100(2), 506–10.
- Rajan, Uday, Amit Seru, and Vikrant Vig, “The failure of models that predict failure: Distance, incentives, and defaults,” Journal of Financial Economics, 2015, 115(2), 237–260.
- Reinhart, Carmen M and Kenneth S Rogoff, This time is different: Eight centuries of financial folly, princeton university press, 2009.
- Rossi, Alberto G, “Predicting stock market returns with machine learning,” 2018.
- Stulz, René M, “FinTech, BigTech, and the Future of Banks,” NBER Working Paper, 2019.
- Sullivan, Tara and James Vickery, “A Look at Bank Loan Performance,” Federal Reserve Bank of New York Liberty Street Economics, 2013.
- Syverson, Chad, “What determines productivity?,” Journal of Economic literature, 2011, 49(2), 326–65.

### Figures (captions and notes)
- Figure 1: NPLs over Assets by pre-GFC IT adoption
  - Caption note: This Figure plots the median share of NPLs over assets for high and low IT adopters. “High IT adoption" is the median share of NPLs over assets for banks with I T_b above the 75th percentile.  “Low IT adoption" is the median share of NPLs over assets for banks with I T_b below the 25th percentile.  We include only banks for which we have regulatory data for at least 14 years.  See subsection 4.1 and section 3 for more details.
- Figure 2: Time-varying Effect of IT adoption on NPLs
  - Caption note: This Figure plots the coefficient and the 95% and 99% confidence intervals of β_τ from the following estimated equation: N P L_b,t = α_b + δ_t + ∑_{τ≠2006} β_τ I T_b · 1[t=τ] + ε_b,t where b is a bank (BHC), t one year between 1996 and 2014, α_b are bank fixed effects, and δ_t are year fixed effects. The dependent variable N P L_b,t is the share of NPLs over assets in b’s regulatory filing for year t. I T_b is the pre-crisis IT-adoption of bank b estimated as described in section 3. The coefficient of 2006 is normalized to zero. Confidence intervals are based on double-clustered standard errors at the bank and year level. See subsection 4.1 and section 3 for more details.
- Figure 3: Loans over pre-crisis Assets by pre-GFC IT-adoption
  - Caption note: This Figure plots the median share of total loans scaled by average pre-crisis (2001-2006) assets for high and low IT adopters. “High IT adoption" is the median share of Loan over pre-crisis assets for banks with I T_b above the 75th percentile.  “Low IT adoption" is the median share of Loan over pre-crisis assets for banks with I T_b below the 25th percentile.  We include only banks for which we have regulatory data for at least 14 years. See subsection 5.3 and section 3 for more details.

### Tables (captions and key model specifications)
- Table 1: Panel Regressions
  - Dependent Variable: NPLs during GFC
  - Selected coefficients and standard errors (double-clustered on bank and year):
    - IT-adoption: -0.0230 (0.017); -0.0280 (0.018)
    - IT-adoption×crisis: -0.160** (0.063); -0.168** (0.065); -0.157** (0.066); -0.170** (0.068); -0.158** (0.063); -0.139** (0.057); -0.131** (0.055); -0.131** (0.056); -0.131** (0.056); -0.131** (0.056); -0.131** (0.056); -0.143** (0.063)
    - Loans×crisis: 0.0201*** (0.006); 0.0199*** (0.006); 0.0177*** (0.006); 0.0175*** (0.005); 0.0186*** (0.006); 0.0189*** (0.006); 0.0189*** (0.006); 0.0192*** (0.006)
    - HP Exposure×crisis: 0.0191** (0.009); 0.0172* (0.008); 0.0172* (0.008); 0.0172* (0.008); 0.0171* (0.008); 0.0171* (0.009); 0.0174* (0.009)
    - Size×crisis: 0.128** (0.052); 0.126** (0.050); 0.115** (0.050); 0.123** (0.055); 0.124* (0.062); 0.124* (0.062)
    - Capital×crisis: -0.000672 (0.005); 0.000698 (0.005); 0.000933 (0.005); 0.000805 (0.005); 0.00113 (0.005)
    - Wholesale×crisis: 0.00831 (0.007); 0.00779 (0.007); 0.00779 (0.007); 0.00762 (0.007)
    - ROA×crisis: -0.0204 (0.048); -0.0206 (0.048); -0.0211 (0.049)
    - Log Wage×crisis: -0.0108 (0.125); -0.00388 (0.124)
    - IT of local competitors×crisis: 0.0475 (0.046)
  - (Within) R-squared: 0.112; 0.140; 0.0111; 0.00997; 0.0363; 0.0553; 0.0613; 0.0611; 0.0621; 0.0620; 0.0617; 0.0623
  - N: 4608 in each column
  - Bank FE: No/Yes per column; Year FE: No/Yes per column
  - Model estimated equation: N P L_b,t = α_b + δ_t + β I T_b · crisis_t + (X_b · crisis_t)' γ + ε_b,t where crisis_t is a dummy for years 2007 to 2010. Bank-level controls X_b defined as averages over 2001-2006 include loans to assets ratio, capital to assets ratio, wholesale funding ratio, ROA, (log) average wages in thousands of USD, (log) assets size in thousands of USD, IT-adoption of local competitors, and HP Exposure. Standard errors double-clustered on bank and year. *p<0.1, **p<0.05, ***p<0.01.
- Table 2: Cross-Sectional Regressions
  - Dependent Variable: NPLs during GFC; other dependent variables across columns: Loans pre-GFC, HP Exposure pre-GFC, Size pre-GFC, Capital pre-GFC, Wholesale pre-GFC, ROA pre-GFC, Log Wage pre-GFC, IT of local competitors pre-GFC, NPLs during GFC (with controls)
  - Selected coefficients and standard errors (robust):
    - IT-adoption: -0.183*** (0.061); -0.648 (0.700); -0.896 (0.664); -0.0931 (0.057); -0.195 (0.420); -0.0459 (0.372); -0.0282 (0.049); -0.0227 (0.018); 0.275*** (0.083); -0.157*** (0.058)
    - Loans: 0.0305*** (0.004)
    - HP Exposure: 0.0242*** (0.005)
    - Wholesale: 0.0251*** (0.008)
    - IT of local competitors: 0.0773 (0.047)
  - R-squared values: 0.0262; 0.0022; 0.0055; 0.00712; 0.000427; 0.0000383; 0.00107; 0.00414; 0.07500; 0.243
  - N: 337 in each column
  - Sample means and Std.Dev. reported:
    - Mean (first column) 1.546; Std.Dev. 1.131
    - Mean (second column) 2.6915; Std.Dev. 3.812
    - Mean (third column) 5.8313; Std.Dev. 2.061
    - Mean (fourth column) 3.913; Std.Dev. 1.19
    - Mean (fifth column) 3.0215; Std.Dev. .437
    - Mean (sixth column) .922; Std.Dev. .41
    - Mean (seventh column) .554; Std.Dev. .86
    - Mean (eighth column) .84; Std.Dev. .351
    - Mean (ninth column) 1.54; Std.Dev. 1.13
  - Estimation equation: Y_b = α + β I T_b + ε_b where Y_b is either average NPLs over 2007 to 2010 or one of the X_b variables averaged over 2001-2006; column (10) includes X_b as controls. Robust standard errors reported. *p<0.1, **p<0.05, ***p<0.01.
- Table 3: Loan-Level Regressions
  - Dependent Variable: Delinquency during GFC (Share of months with past due>90 days; Ever past due>90 days)
  - Selected coefficients and standard errors:
    - IT adoption: -0.471** (0.191); -0.459** (0.169); -0.348** (0.145); -0.323** (0.118); -0.106** (0.041); -0.0377** (0.016)
    - FICO score: -2.578*** (0.284); -1.125*** (0.181); -1.042*** (0.088); -0.258*** (0.036)
    - DTI: 0.565*** (0.052); 0.248*** (0.022); 0.246*** (0.019); 0.0994*** (0.012)
    - LTV: 1.075*** (0.129); 0.543*** (0.056); 0.541*** (0.058); 0.185*** (0.006)
    - IT adoption×Low FICO: -0.198*** (0.064)
    - IT adoption×High FICO: -0.00732 (0.029)
  - Estimation methods: OLS (columns 1–6) and Probit (column 7)
  - Fixed effects: Organization Year FE (No/Yes across columns); Postal Code FE (No/Yes across columns as noted)
  - N: 3,451,671 observations in each column
  - Dependent variable means and Std.Dev.:
    - Mean (columns 1–6) 3.44; Std.Dev. 14.321
    - Mean (column 7) 1.5; Std.Dev. 15.1215
  - Model estimated equation: Delinquent_l = α_{z(l)} + δ_{o(l)} + β I T_{b(l)} + X'_l γ + η_l where l is a mortgage held by Freddie Mac and originated before 2007, α_{z(l)} are 3-digit postal code fixed effects, δ_{o(l)} are origination year fixed effects, I T_{b(l)} is pre-crisis IT-adoption of the bank which sold the mortgage to Freddie Mac (22 banks). Dependent variable is either share of months between 2007 and 2010 with delinquency (> 90 days past due) or a dummy indicating ever delinquent between 2007 and 2010. Both multiplied by 100 except in column (7). Controls X_l include FICO score, debt servicing to Income (DTI), and Loan-to-Value (LTV) at origination. All independent variables are standardized to have a mean of 0 and a standard deviation of [text truncated in source].

*Italic: Content from "References" section of wpiea2020014-print-pdf*

### 1. Column (1) excludes all fixed effects and controls, column (2) excludesX

### wpiea2020014-print-pdf - 1. Column (1) excludes all fixed effects and controls, column (2) excludesX

### Table 4: NPLs, IT adoption, and Executives’ “tech-orientation”
- Dependent Variable variants: NPLs during GFC; IT-adoption during GFC.
- Estimations reported: OLS (columns 1–3), IV (column 4).
- Key coefficients and standard errors (in parentheses):
  - IT-adoption: -0.138* (0.076) in column (1); -1.719* (1.044) in column (3).
  - Executives’ “tech-orientation”: -0.155*** (0.047) in column (1); 0.0900* (0.051) in column (3).
- R-squared: 0.0141 (col 1); 0.0210 (col 2); 0.00967 (col 3).
- Sample size: N 249 for all reported columns.
- Notes on specification:
  - Dependent variable Y_b is ratio of NPLs to assets averaged between 2007 and 2010 (columns 1, 2, and 4) or pre-crisis IT adoption (column 4).
  - Independent variable X_b is either pre-crisis IT adoption or average “tech-orientation” of top executives (CEOs, CFOs, Presidents).
  - “Tech-orientation” computed as total “tech-related” keywords divided by total words in biographies.
  - Column (4): IT-adoption instrumented with executives’ “tech-orientation”; column (3) is first stage for (4).
  - Sample kept constant by dropping observations with missing values. Robust standard errors in parentheses. *p<0.1, **p<0.05, ***p<0.01

### Table 5: Executives’ “tech-orientation” and Compensation
- Dependent variables: NPLs (columns 1–2), IT-adoption (columns 3–4).
- Key coefficients and standard errors (in parentheses):
  - Executives’ “tech-orientation”: -0.173*** (0.062) in column (1); -0.168*** (0.062) in column (2); 0.104* (0.057) in columns (3) and (4).
  - Log Compensation: -0.0375 (0.060) in column (2); -0.00208 (0.053) in column (4).
- R-squared: 0.0226 (col 1); 0.0244 (col 2); 0.0136 (cols 3 and 4).
- Sample sizes: N 237 (cols 1–2); N 149 (cols 3–4).
- Notes on specification:
  - Equation: Y_b = α + β ExecutiveIT_b + γ Comp_b + ε_b.
  - Comp_b is log of average total compensation earned by top executives in 2007; excluded in columns (1) and (3).
  - Sample kept constant by dropping observations with missing values. Robust standard errors reported. *p<0.1, **p<0.05, ***p<0.01

### Table 6: Lending Regressions
- Dependent Variable: Loan Growth (crisis) — loan growth over assets averaged over 2007 to 2010.
- Key coefficients and standard errors (in parentheses):
  - NPLs during the GFC: -0.926*** (0.159) in column (1); -1.030*** (0.187) in column (3).
  - IT-adoption: 0.378** (0.182) in column (1); 0.331* (0.196) in column (3).
- R-squared: 0.0127 (col 1); 0.0928 (col 2); 0.0961 (col 3); 0.175 (col 4).
- Sample sizes: N 343 (cols 1 and 3); N 336 (cols 2 and 4).
- Controls: reported as No (cols 1 and 3) and Yes (cols 2 and 4).
- Notes on specification:
  - Equation: ΔLoans_GFC_b = α + β X_b + ε_b where X_b is pre-crisis IT-adoption or share of NPLs over assets (averaged 2007–2010).
  - Robust standard errors in parentheses. *p<0.1, **p<0.05, ***p<0.01

### Appendix — Figures (A1–A4)
- Figure A1:
  - Cross-sectional distribution of ratio of NPLs to assets averaged 2007–2010 (top panel) and pre-crisis IT adoption I T_b.
- Figure A2:
  - Median share of NPLs over assets for high, medium, and low IT adopters.
  - “High IT adoption”: I T_b above 75th percentile; “Low IT adoption”: I T_b below 25th percentile; “Median IT adoption”: I T_b between 25th and 75th percentiles.
  - Includes banks with at least 14 years of regulatory data.
- Figure A3:
  - Median share of NPLs scaled by average pre-crisis (2001-2006) assets for high and low IT adopters.
  - Definitions of high/low as above; includes banks with at least 14 years of regulatory data.
- Figure A4:
  - Robustness of executives’ results to changes in keywords list.
  - For each word used to define technology orientation, a new measure is created leaving that word out; plot shows coefficient of columns (2)-(4) of Table 4 for these alternative measures.
  - Dashed line reflects estimates of columns (2) to (4) of Table 4.

### Table A1: Robustness of Main Panel Regression
- Dependent Variable: NPLs.
- Coefficients for IT-adoption × crisis and standard errors (in parentheses):
  - Column (1) Baseline: -0.165** (0.068)
  - Column (2) PCs per Emp: -0.243* (0.120)
  - Column (3) HW I T winsorized: -0.158** (0.069)
  - Column (4) HW NPLs winsorized: -0.161** (0.063)
  - Column (5) Loans Broad def.: -0.242** (0.095)
  - Column (6) As of 2006: -0.214** (0.080)
  - Column (7) Bank Clustering: -0.380* (0.183)
  - Column (8) (alternative clustering): -0.165*** (0.051)
- R-squared values: 0.00944 (col 1); 0.00376 (col 2); 0.00794 (col 3); 0.0108 (col 4); 0.00867 (col 5); 0.00993 (col 6); 0.00530 (col 7); 0.00944 (col 8).
- N observations: reported per column as 469, 250, 354, 692, 469, 246, 924, 692 (as listed).
- Fixed effects: Bank FE Yes; Year FE Yes for all columns.
- Notes on specification:
  - Equation: NPL_b,t = α_b + δ_t + β IT_b · crisis + ε_b,t with crisis_t indicating years 2007 to 2010.
  - IT_b is pre-crisis IT-adoption. Column-specific modifications: column (2) uses PCs per employee; column (3) measure winsorized at 5 percent; column (4) NPLs winsorized; column (5) NPLs normalized by loans; column (6) broader NPLs definition; column (7) NPLs normalized by average pre-crisis assets; column (8) standard errors clustered only at bank level.
  - Standard errors double-clustered on bank and year for columns (1)-(7). *p<0.1, **p<0.05, ***p<0.01

### Table A2: Robustness of Executives’ “tech-orientation” to the inclusion of Non-Base Compensation
- Dependent variables: NPLs (columns 1–2) and IT-adoption (columns 3–4).
- Key coefficients and standard errors (in parentheses):
  - Executives’ “tech-orientation”: -0.477*** (0.151) in column (1); -0.475*** (0.149) in column (2); 0.212* (0.114) in column (3); 0.213* (0.115) in column (4).
  - Non-Base Compensation: -0.281 (0.520) in column (2); -0.191 (0.665) in column (4).
- R-squared: 0.0844 (col 1); 0.0875 (col 2); 0.0241 (col 3); 0.0262 (col 4).
- Sample sizes: N 80 (cols 1–2); N 79 (cols 3–4).
- Notes on specification:
  - Equation: Y_b = α + β ExecutiveIT_b + γ NonBaseComp_b + ε_b where NonBaseComp_b is average share of non-base compensation over total compensation for top executives in 2007.
  - NonBaseComp_b excluded in columns (1) and (3).
  - Sample kept constant by dropping observations with missing values. Robust standard errors in parentheses. *p<0.1, **p<0.05, ***p<0.01

*Content reproduced from the specified PDF chapter/section.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020014-print-pdf.pdf_