Bank Relationships and the Geography of PPP Lending

The PPP reopened on January 11, 2021 after a third round of funding was provided. However, this paper focuses on lending done in the first two rounds.

The PPP FAQs put out by the Treasury and SBA clarified that existing customers who were previously verified by the lender would not need to be reverified for compliance with the Customer Due Diligence rule or Bank Secrecy Act. See answers to questions 18 and 25 here: https://​home.​treasury.​gov/​system/​files/​136/​Paycheck-Protection-Program-Frequently-Asked-Questions.​pdf.

Bartik et al. (2020) use survey data to show that banks are more likely to approve loans to relationship borrowers. Amiram and Rabetti (2020); Duchin et al. (2022) use data from public firms in the PPP to document that relationship borrowers are more likely to get credit and get credit faster.

Erel and Liebersohn (2020) show that fintech lenders have a higher market share in ZIP codes with fewer branches or a larger minority share of the population. Li and Strahan (2021) show that banks allocate more PPP loans to counties where they did more small business lending before COVID. Granja et al. (2022) show that ZIP codes with more active banks had higher levels of PPP lending, though employment effects were modest. Faulkender et al. (2023) show that counties served more by community banks received credit earlier and experienced a smaller rise in unemployment.

Technology is interpreted broadly as any factor that affects a bank’s willingness or ability to quickly extend credit. Complementary papers provide insight into how factors such as previous SBA experience (Granja et al. 2022), concerns about regulatory fallout (Faulkender et al. 2023), or the suite of technological products utilized by the bank (Pogach and Kutzbach 2022) contribute to such technological differences.

This data is available here: https://​www.​geocod.​io/​geocoded-ppp-loan-data/​.

Some lender names in the PPP data can match to multiple banks. I disambiguate by selecting fuzzy name matches that also match on ZIP code, then state, then finally selecting the prospective match with the highest PPP lending volume.

Distance is calculated using the haversine formula. The minimum distance is found using the BallTree module from scikit-learn.

Small business loans are C&I loans with an original balance under $1 million dollars.

I use the data on private jobs in 2017, as this is the most recent year that reports employment disaggregated by firm size. Alaska is missing for 2017, so I instead use 2016 data there.

Earnings variables include the share of jobs in the tract earning $1250 or less per month, the share of jobs earning over $3333 per month, and the share of jobs by workers with at least a bachelor’s degree. Since PPP loan sizes are pinned to monthly payroll, these variables along with the number of employees in PPP-eligible firms are important controls in specifications predicting PPP lending.

I use data from 2019 in order to reflect prepandemic small business lending relationships. The data is only reported by banks with at least $1.3 billion in assets, and thus does not reflect the extent of small business lending relationships with the smallest banks.

Tract locations are measured by the centroids in the 2010 Census Gazetteer files.

Outcomes considered include: PPP loan volumes, average days until origination, and the shares of PPP lending accounted for by fintech lenders or local lenders.

Brevoort and Wolken (2009) document that the median distance between small firms and the bank servicing them is about 3 miles, with nearly 90% of the lenders being within 30 miles.

Binscatters are constructed using the Binsreg command in Stata (Cattaneo et al. 2019), controlling for loan size, a small firm indicator, census tract characteristics (an urban indicator, the share of employees that are nonwhite, and the logarithms of prepandemic small business employment and small business lending), and county fixed effects.

All specifications include county and three-digit NAICS fixed effects, and control for loan size, whether the business has fewer than five employees, and tract characteristics (an urban indicator, the share of employees that are nonwhite, and the logarithms of prepandemic small business employment and small business lending). Regressions are either unweighted, or weight by the inverse of the number of PPP originations at the bank.

Since borrowers choose the nonrelationship lender with the best processing time, \(T_b^N\) is equalized for all banks that do nonrelationship lending. Relationship prioritization results in \(T_b^R<T_b^N\), so borrowers would never switch from a bank that does nonrelationship lending.

\(T_{b}^R\) is indeterminate if \(\rho _b=1\), taking values in the range \([\frac{T}{1+\gamma },T]\), reflecting the range of \(\tau _b\) that is low enough to serve all relationship customers, but high enough to not attract nonrelationship customers after. Since there are no observations in the data with this intensity, it does not matter for the estimation.

The second line comes from the fact that \(f(d_{i,b}|\rho _b)=\rho _b^{-1}f^R(d)+(1-\rho _b^{-1})f^N(d)\), \(f(d_{i,b}|R_{i,b}=1,\rho _b)=f^R(d)\) and \(P(R_{i,b}=1|\rho _b)=\rho _b^{-1}\) when \(\rho _b>1\).

Over 20% of businesses with fewer than five employees were unaware of the program as of the time the initial funding was exhausted, whereas the vast majority of larger businesses were aware of it within a couple days of opening (Humphries et al. 2020).

SBA data is available here: https://​data.​sba.​gov/​dataset/​7-a-504-foia.

This result is consistent with the model; clients at lower intensity banks are willing to switch lenders at higher switching costs, resulting in higher average wait times.

I consider the upper range here to reflect the 95th percentile of PPP intensity, whereas in Table 2 the upper range is determined by the highest PPP intensity in the sample. I do this because the predictions from the regression specification assume effects are linear in ln(PPP Intensity), making extrapolation to extreme values more problematic than in the model where effects are concave in the right tail.

The positive coefficient in the unweighted specification without the lending bank fixed effect is driven by the top four banks, which predominantly served nearby clients and were slow to extend credit.

Additionally, Appendix Table 7 presents similar estimates, but measuring local supply conditions with the share of deposits in nearby branches that are from the top four banks. Given the geographic extent of these banks’ operations, there is less concern about reverse causality for this supply measure.

The PPP program essentially provided credit for up to 2.5 times monthly payroll for businesses with 500 or fewer employees. This means that the logarithm of total PPP lending in an area will roughly be the sum of: i. ln(Small Business Employment), ii. ln(2.5 \(\times \) Average Monthly Payroll of Eligible Firms) and iii. ln(Share of Eligible Payroll Funded). I control for the logarithm of employment in firms with fewer than 500 employees (to account for i.) and the following variables pertaining to average earnings (to account for ii.): the share of jobs earning under $1250 per month, the share of jobs earning over $3333 per month, and the share of workers with a college degree. Effects of other variables can thus be more reasonably attributed to bank supply conditions.

In Appendix Table 5, the coefficient on the interaction between ln(PPP Intensity)\(_b\) and Local Branch\(_{i,b}\) is -2.5 in the unweighted specification and -0.6 in the weighted specification.

The fintech share is the share of loans made by the following lenders: Celtic Bank, Cross River Bank, Intuit, Fundbox, Kabbage, Readycap and WebBank. This includes both nonbanks, and some prominent banks with significant lending through fintech partnerships. Fintechs make up about a quarter of the decline in market share of local banks when market shares are measured by counts rather that volumes, as they tended to originate smaller loans.

The relationship between racial composition and loan volumes is unclear. One specification shows notable effects for the nonwhite share and the other notable effects for the college share. Assessing differences in the ability to get PPP loans is complicated by the racial wage gap: lower PPP lending could reflect lower average earnings among minority employees (reducing loan sizes) or greater frictions in accessing PPP credit (reducing loan counts).

Race is not reliably reported in the PPP data, so distinguishing these channels is better left to work utilizing other sources to identify race. Such work shows that black-owned firms are less likely to receive PPP funds, and when they do receive funds, it is less likely to be from banks (Howell et al. 2021; Chernenko and Scharfstein 2022). These differences appear to predominantly driven by disparities in applications rather than approvals, with bank relationships accounting for a bit less than a quarter of the disparity (Chernenko et al. 2023).

\(X_j\) is as in Table 4, but with additional dummy variables indicating whether there are no banks within a particular distance band. To maintain the same sample as in the previous analysis, ln(PPP intensity) is recoded to zero when there are no banks within a band and tracts are dropped if there are no banks within two miles.