Walking the walk? Bank ESG disclosures and home mortgage lending

Abstract

We show that banks with high environmental, social, and governance (ESG) ratings issue fewer mortgages in poor localities—in number and dollar amount—than banks with low ESG ratings. This lending disparity happens at both the county and census tract level, worsens in disaster areas of severe hurricane strikes, is robust to alternative ESG ratings (including using only the social (S) component), and cannot be explained by banks’ differential deposit networks. We find no difference in mortgage default rates between high- and low-ESG banks, rejecting an alternative explanation based on differential credit screening quality. We report a complementary, not substitution, relation between high-ESG banks’ mortgage lending and their community development investments (like affordable housing projects) in poor localities. Loan-application-level analyses confirm that high-ESG banks are more likely than low-ESG banks to reject mortgage loans in poor neighborhoods. The evidence hints at social wash: banks deploy prosocial rhetoric and symbolic actions while not lending much in disadvantaged communities, the social function they arguably ought to perform. Community Reinvestment Act (CRA) examinations partially undo the social wash effect.

1 Introduction

The notion that companies should behave in socially responsible ways beyond maximizing profit, once dismissed by many scholars and business leaders as untenable (Pigou 1920; Friedman 1963, 1970), is now popular. The number and scope of corporate claims about environmental, social, and governance (ESG) performance have exploded recently. At the start of 2020, about 90% of firms in the S&P 500 index disclosed ESG data, and more than $17 trillion was invested in funds that select portfolio firms using explicit ESG criteria.¹ Many asset managers and data companies produce ESG ratings, often relying on firms’ own disclosures for ESG-related data. However, some regulators question the truthfulness of these ESG disclosures, arguing that they misrepresent companies’ actual social or environmental practices (Peirce 2019; Roisman 2020). In their view, firms talk the ESG talk, but often do not walk the ESG walk.

We study corporate social responsibility in commercial banks whose main business role—providing credit to local communities—can create widespread benefits. We study whether banks with high ESG ratings originate more home-purchase mortgage loans in poor localities than banks with low ESG ratings. Our inquiry is motivated by the long-standing principle that banks are socially obliged to expand home mortgage access in underserved localities while maintaining safe and sound operations (Community Reinvestment Act (CRA) of 1977; Bernanke 2007).² However, credit market failures and discrimination have historically impeded lending to otherwise creditworthy borrowers in poor neighborhoods (Lang and Nakamura 1993; Barr 2005; Rothstein 2017). Given the socioeconomic benefits of homeownership (Dietz and Haurin 2003), the interplay of banks’ ESG and mortgage credit provision, especially in poor localities, merits a close look.

Our sample combines ESG data from Refinitiv with home-mortgage-lending data released by the Federal Financial Institutions Examination Council (FFIEC) under the Home Mortgage Disclosure Act (HMDA) from 2002 to 2018. To separate banks’ mortgage supply from the possibly confounding effects of local mortgage demand, our empirical identification focuses on the within-locality-year variation in mortgage lending. Our rationale is that banks with differing ESG ratings face a similar mortgage demand for properties in the same locality in a given year, as evidenced by banks in a given locality receiving similar loan applications.

We have two main findings. First, high-ESG banks originate fewer home purchase loans—in both number and dollar amount—annually in poor counties than low-ESG banks do. Second, within the same county, high-ESG banks are less likely than low-ESG banks to lend on properties in poor census tracts that should be served by both types of banks under the CRA. These findings are not driven by bank size, since the correlation between bank ESG ratings and bank size is low and since excluding the largest banks or multinational banks does not change the results much. The results hold using only the social (S) component of ESG ratings, are not driven by differential underwriting standards between high- and low-ESG banks, and persist across alternative ESG measures from Bloomberg, MSCI/KLD, and S&P Global.

To assess causality, we instrument a bank’s ESG rating with the voting percentage for the Democratic presidential candidate in the bank’s headquarter state in the last presidential election. This instrument relies on firms in Democratic states being more likely than firms in Republican states to promote ESG practices (Di Giuli and Kostovetsky 2014). Moreover, states’ election outcomes vary over time, providing plausibly exogenous time-series variation to banks’ ESG pressure. Using a two-stage least squares (2SLS) framework, we show that increases in banks’ ESG ratings correspond to significant declines in the banks’ mortgage lending in poor neighborhoods, supporting our main inference.

We propose a social wash effect, in which firms selectively undertake and advertise symbolic prosocial activities in their ESG disclosures to divert attention from their weak lending records in poor areas.³ Several tests corroborate this effect. First, we use severe hurricanes as a setting. After hurricanes, mortgage demand rises in disaster areas as residents seek to rebuild destroyed properties, but banks’ social wash incentives also increase. Banks might avoid lending in poor disaster areas because the properties there often lack flood insurance, which impedes recovery, depresses local home values, and increases mortgage default risk. Meanwhile, active media attention and news coverage around severe hurricanes such as Katrina likely prod banks to profess social altruism. We show that high-ESG banks are more likely than low-ESG banks to halt their mortgage lending in low-income counties hit by big hurricanes.

Second, we test whether CRA rules and enforcement mitigate the ESG–mortgage lending incongruence. Under the CRA, banks are examined periodically on their lending in low-income communities. Banks failing the exam are prohibited from opening new branches or engaging in mergers and acquisitions until their lending record improves. We report that high-ESG banks’ lending curtailment in low-income areas is mitigated, though not eliminated, among the banks that rated highest in recent CRA exams. High-ESG banks also increase mortgage lending more quickly in poor localities after CRA rating downgrades. The collective evidence suggests that CRA enforcement partially undoes banks’ social wash behavior.

Third, we use detailed HMDA loan application data to unpack how banks’ ESG ratings covary with individual mortgage application decisions. We find that high-ESG banks are less likely than low-ESG banks to approve home mortgage loans in poor areas, after controlling for borrower attributes like income, debt-to-income ratio, race, and ethnicity. However, high-ESG banks are just as likely as their low-ESG counterparts to offer high-yield loans. The pricing result mitigates the concern that high-ESG banks’ poor lending record in poor areas is driven by innately risky borrowers selecting into high-ESG banks; if that were the case, we would expect high-ESG banks to price in the risk with higher interest rates.

Our study makes several contributions. First, our results suggest that bank managers promoting ESG may not put their money where their mouths are. These managers can back up their ESG credentials by, for example, lending in poor localities. Second, our findings indicate that ESG ratings are noisy. Ratings providers should tailor their scoring to sustainability issues that are material to specific firms and industries or, at the very least, incorporate the Sustainability Accounting Standards Board (SASB) materiality frameworks. Third, we show that current regulatory regimes, including CRA enforcement, do not fully curb disparate mortgage lending. Fourth, we advance a keen debate on the role of ESG practices and disclosures (e.g., Moser and Martin 2012; Khan et al. 2016; Christensen et al. 2019; Grewal and Serafeim 2020; Larcker and Watts 2020; Raghunandan and Rajgopal 2021a, 2021b).

2 Institutional background and hypotheses

As of December 31, 2019, total U.S. mortgage balances ($9.56 trillion) constituted 68% of total household debt ($14.15 trillion).⁴ Banks enhance the economic health of communities by meeting their housing credit needs. Banks have been criticized for historically refusing or reducing credit access to minority or poor households and neighborhoods. The term “redlining” refers to bankers’ practice of marking off neighborhoods (often in red lines on maps) as “hazardous” for lending due to their geographical location and racial composition (Benston 1981; Holmes and Horvitz 1994).

To counter discriminatory lending practices, Congress enacted a series of laws aimed at equalizing access to home mortgages, including the Fair Housing Act of 1968 (FHA), the Equal Credit Opportunity Act of 1974 (ECOA, 15 U.S.C. § 1691), the Home Mortgage Disclosure Act of 1975 (HMDA, 12 U.S.C. § 2801), and the Community Reinvestment Act of 1977 (CRA, 12 U.S.C. § 2901–2908). FHA and ECOA address racial discrimination in mortgage lending, while the CRA and the HMDA more broadly address under-lending in low- and moderate-income communities. When enacting the CRA, Congress stated that financial institutions have a “continuing and affirmative” obligation to help meet the credit needs of low- and moderate-income areas in which banks are chartered (12 U.S.C. §2901). The CRA’s goal is to induce lenders to look harder for profitable lending opportunities in disadvantaged neighborhoods that they otherwise would avoid, within the bounds of safety and soundness standards.

Although banks have substantially reduced discriminatory lending and information barriers in historically underserved markets, disparities in mortgage access persist.⁵ Applicants from poor neighborhoods and minority applicants (who tend to inhabit these areas) are more likely to be denied a mortgage loan and charged higher interest rates, even after controlling for relevant economic attributes (e.g., Munnell et al. 1996; Ladd 1998; Bartlett et al. 2022).⁶ The underserved and underbanked segment is especially vulnerable to predatory lending (e.g., Morgan 2007), so access to traditional banking credit may protect them. The banking sector, as a quid pro quo for governmental subsidies like federal deposit insurance and federal backstop, is arguably obliged to reduce structural inequalities and disparate lending in poor neighborhoods (Barr 2005; Bernanke 2007). Homeownership has long been viewed as a crucial route for low-income households to build wealth and move up the social ladder. Thus, any discussion or metric of banks’ ESG (even if “ESG” is just an umbrella term) must consider the banks’ mortgage lending to underserved home buyers, just as emitting less carbon dioxide is atop oil-and-gas firms’ ESG agenda.

2.1 Social wash hypothesis

Whether banks that profess strong ESG performance lend more in poor neighborhoods is theoretically unclear. Banks can selectively undertake and promote symbolic ESG actions to present a socially responsible public image but avoid tangible actions that generate real social benefits. Rather than meet housing credit needs in poor areas, banks with high ESG scores may pay lip service to an ESG agenda. We refer to this as the social wash hypothesis.

Some firms engage in social wash because it benefits them, at least in the short term. Firms market themselves as socially or environmentally conscious rather than profit maximizing because the public has become increasingly enthusiastic about corporate goodness (Mishra and Modi 2016; Dyck et al. 2019). By disclosing prosocial efforts while obscuring antisocial ones, firms can placate public opinion and forestall activists (Dhaliwal et al. 2011; Grewal et al. 2019). Some consumers pay premium prices for products and services from socially conscious firms (Bhattacharya and Sen 2003; Berens et al. 2005; Servaes and Tamayo 2013). Wall Street also incentivizes social wash. Asset managers’ fervor for socially responsible investments pressures portfolio companies to adopt ESG strategies (Friedman and Heinle 2016; Chen et al. 2020).

Whether to social wash depends on a bank’s cost–benefit calculations. For example, hiring public-relations specialists and/or purchasing ESG advisory services may be less costly to large firms than to small ones. Banks that already finance mortgages locally likely see less need to flaunt their social achievements; overclaiming ESG could backfire in areas where residents know and trust their bankers.⁷ In theory, the marginal benefits of social wash are less when it is easier to detect and when the punishments for it are harsher. Firms caught engaging in social wash could (but rarely do) suffer reputational loss, class-action lawsuits, regulatory interventions, or expulsion from ESG indices, which could, in turn, reduce the firms’ stock prices (Hoffman 2013; Lyon and Montgomery 2015).

Organizational behavior theories posit that superficial conformity with social norms disconnects firms’ real actions from their external facades (e.g., Meyer and Rowan 1977; Cho et al. 2015). The lack of standardization, transparency, and enforcement for ESG disclosures (Christensen et al. 2019; Grewal and Serafeim 2020) offers managers flexible reporting, increasing this disconnect. ESG data are seldom audited, and many companies lack adequate internal controls to limit ESG misstatements. ESG providers are conflicted insofar as they sell ESG consulting services to the companies they rate, eroding ratings integrity (Eaglesham 2022).⁸ While social wash is believed to be pervasive (Raghunandan and Rajgopal 2021a, 2021b), banks’ social wash likely misleads local communities and households about mortgage credit access and prolongs poverty.

Banks can raise their ESG metrics by telling rosy stories about policies, products, and performance (see Lyon and Montgomery 2015); adding pleasing photos in annual reports, websites, and social media (picture a family of four smiling happily in their new home); sponsoring charity programs; selectively disclosing favorable data; and, more recently, underwriting or issuing ESG-labelled instruments like green bonds.⁹ We do not argue that all banks must lend in low-income communities. There are rational reasons why some banks lend less than others in these areas—profit maximization, deposit network, underwriting standards, etc. The social wash hypothesis predicts that banks claim more than they deliver in the social arena.¹⁰

We formulate the social wash hypothesis as follows, in alternative form:

H1a: Banks with high ESG ratings issue fewer mortgage loans in poor neighborhoods than banks with low ESG ratings (social wash hypothesis).

However, other reasons could yield the same prediction. One reason is that the marginal benefits of undertaking costly yet socially desirable activities to boost ESG credentials are higher when banks have low ESG ratings, which would lead low-ESG banks to lend more in poor communities. Another possibility is that high-ESG banks are more prudent underwriters and rationally reduce lending in low-income communities to mitigate default risk. Banks could substitute mortgage origination in low-income areas with investment programs designed to promote affordable housing and local economic development or lend in less developed countries.¹¹ We evaluate these other arguments in Section 5 and Online Appendix.

2.2 Social signaling hypothesis

Alternatively, if banks accurately disclose their social and environmental actions, then banks with high ESG performance will likely provide more home mortgages in low-income areas than banks with low ESG performance. Under signaling theory, in markets with asymmetric information, high performers will reveal credible data about their performance to try to separate themselves from low performers (Lys et al. 2015). In a separating equilibrium, high-ESG banks will publicize their superior ESG performance, while low-ESG banks will not disclose.

According to Bénabou and Tirole (2010), self-signaling plays an important role in driving both individuals’ and firms’ decisions to undertake prosocial behavior and make sure people know about it. The equilibrium is one in which banks’ ESG disclosures faithfully represent their true commitment, i.e., the banks with better home mortgage lending in poor neighborhoods are those that achieve higher ESG scores. If a company consistently overstates its corporate social responsibility achievements, market participants will discern such behavior and the signal will lose its credibility. We label this the social signaling hypothesis.

The social signaling hypothesis does not require that bank ESG metrics precisely report each of a bank’s multipronged social practices. Mortgage lending behavior is a function of various factors, including the bank’s business model and strategic operations, so it cannot be reduced to a single metric. The hypothesis simply requires that social and community engagements be positively correlated; a bank that does well in one social dimension is expected to do well in others, given the bank’s overarching commitment to society. We formulate the social signaling hypothesis as follows:

H1b: Banks with high ESG scores issue more mortgage loans in poor neighborhoods than banks with low ESG scores (social signaling hypothesis).

3 Sample selection and empirical design

3.1 Sample selection

Table 1 details the sample selection. We obtain ESG data from Refinitiv (previously issued by Thomson Reuters), which provides standardized ESG scores for listed companies since 2002. Firms’ ESG performance is evaluated across three main dimensions (“pillars”)—environmental, social, and governance—using public data sources such as the firms’ annual reports and corporate social responsibility reports, the news media, and the websites of nongovernment organizations (NGOs). Each pillar has a few subcategories, each of which is assigned a numerical rating based on 178 comparable metrics for the subcategories.¹² Environmental and social ratings are benchmarked to industry averages (Thomson Reuters Business Classifications (TRBC)), and governance ratings are benchmarked to country averages, facilitating direct comparisons among peer firms. Refinitiv also discounts a firm’s overall ESG score for ESG-related controversies (such as business ethics issues or consumer complaints), which reduces the size bias in ESG ratings. We use the controversy-adjusted ESG score in our analyses. We collect ESG ratings for listed bank holding companies, which results in an initial sample of 915 annual ESG observations for 181 bank holding companies over the period 2002–2018.

Table 1 Sample selection and summary statistics

We obtain home mortgage data from HMDA data compiled by FFIEC. Passed by Congress in 1975 and implemented by Regulation C, the HMDA requires mortgage lending institutions to report detailed data about the home mortgage applications they receive, which lets regulators and the public detect possible discriminatory lending practices. The HMDA dataset includes the lending institution name, the application disposition (i.e., acceptance, rejection, withdrawal), the property location, the loan purpose (e.g., home purchase, refinancing), and data about the applicant’s income, race, ethnicity, and gender. We restrict our main analyses to single-family, home-purchase, conventional (i.e., not backed by government agencies like the Federal Housing Administration (FHA)) loans approved by lenders.¹³ We match lending institutions in the HMDA dataset with bank holding companies in the ESG dataset using the regulatory holder identifier. Using this approach, we match 174 out of 181 BHCs.

We aggregate individual loan applications to the bank-county-year level to capture a bank’s mortgage lending share in a county-year. This step leads to a sample of 253,462 bank-county-year observations associated with 174 BHCs (or “banks”). We obtain counties’ annual poverty rates from the U.S. Census Bureau’s Small Area and Income Poverty Estimates (SAIPE) program, and bank financial data from Federal Reserve Y-9C reports. Deleting observations with missing bank financial data reduces the sample to 250,913 bank-county-year observations for 172 BHCs. We retrieve banks’ deposit holdings in a county-year from the FDIC’s Summary of Deposits file and aggregate the data at the holding company level for each county-year. Finally, to identify the within-county-year ESG effect, we require that at least two BHCs lend in each county-year, which yields a final sample of 243,882 bank-county-year observations from 172 BHCs.

We also conduct a bank-tract-year analysis. We first obtain each bank’s CRA assessment areas from the FFIEC CRA Disclosure Flat File Table D6, yielding an initial sample of 3,246,007 bank-tract-year observations associated with 178 banks. Annual poverty rate data for census tracts are taken from the FFIEC Census File. Removing observations with missing bank financial data leads to a sample of 3,221,601 bank-tract-years for 177 banks. Finally, we require each tract-year to have at least two banks making home mortgage loans. Our test sample has 2,978,042 bank-tract-years for 177 BHCs.

3.2 Empirical model

We analyze whether banks with differing ESG ratings vary in their home mortgage lending within a geographical area (county or census tract). We ask if banks with high ESG ratings lend more to homebuyers in a given area-year than banks with low ESG ratings and, more important, if and how this effect varies with an area’s poverty rate. The social wash (social signaling) hypothesis predicts that high-ESG banks will be underrepresented (overrepresented) in low-income neighborhood mortgages. Our empirical tests are in an indirect/reduced form because we do not directly use firms’ disclosures but instead use ESG ratings that, according to prior research (e.g., Cho et al. 2012) and Refinitiv ratings methodology reports, rely heavily on firm ESG disclosures. We begin by estimating a within-county-year regression model:

$$ MGNUMSH{R}_{icy}\ \left( MGAMTSH{R}_{icy}\right)={\beta}_1 ES{G}_{iy}+{\beta}_2 ES{G}_{iy}\times CNTYPOVERT{\mathrm{Y}}_{cy}+{\beta}_3 DEPCNTYSH{R}_{icy}+{\chi}_{iy}+{\alpha}_{cy}+{\lambda}_i+{\epsilon}_{icy}, $$

(1)

where subscripts i, c, and y represent a bank, county, and year, respectively. The unit of observation is the bank-county-year. ESG is a bank’s annual ESG score. The dependent variable is MGNUMSHR (MGAMTSHR), which is a bank’s share of the home-purchase mortgage lending in the county, defined as the number (dollar amount) of home-purchase mortgages extended by a bank in a county-year divided by the total number (amount) of home-purchase mortgages extended in that county-year. CNTYPOVERTY is a county’s annual poverty rate. DEPCNTYSHR represents a bank’s deposit holdings in a county as a percentage of that county’s total deposit holdings across all banks. We include DEPCNTYSHR because the share of a bank’s mortgage lending and its deposit holdings in a county-year are likely correlated.

The vector χ_iy comprises bank-year control variables: BANKSIZE is the natural logarithm of a bank’s total assets (in thousands of dollars); NPL is the ratio of nonperforming loans to total loans; TIER1RAT is the Tier1 risk-based capital ratio; LOANGROWTH is the average annual loan growth over the trailing two years; DEPTOLOAN is total deposits scaled by total loans; LARGETIMEDEP is the ratio of large time deposits (i.e., those with amounts greater than the FDIC deposit insurance coverage limit) to total deposits; COMMERCIAL is the amount of commercial loans (i.e., commercial real estate loans, construction loans, and commercial and industrial loans) divided by total loans, reflecting banks’ commercial lending specialization; and MARKETING is the annual marketing expenses divided by total annual non-interest expenses. We winsorize all bank-year control variables at the 1st and 99th percentiles to mitigate data errors and outliers, except for BANKSIZE (which is in log form).

We include county-year fixed effects, denoted by α_cy. In Online Appendix Table A2, we observe similar applicant and loan attributes (e.g., borrower income, debt-to-income ratio) for mortgage applications received by banks with differing ESG ratings in the same county-year, suggesting that mortgage demand varies little across those banks. Thus, any within-county-year variation should arise from banks’ differential mortgage supply. β₁ captures the effect of a bank’s ESG ratings on its home mortgage lending share in a county-year that has a poverty rate of zero. We focus on the coefficient β₂,which reflects how a bank’s ESG ratings vary with its home mortgage lending, conditional on the county’s poverty rate. A negative (positive) β₂ indicates that high-ESG banks lend less (more) than low-ESG banks in poor counties. We also include bank fixed effects, denoted by λ_i, to control for the possible confounding effects of unobserved bank attributes. Standard errors are double-clustered at the county and bank levels since the residuals are likely correlated along those two dimensions. Our inferences are robust to double clustering by bank and year and double clustering by county and year. We obtain similar results when interacting CNTYPOVERTY with all the control variables in the model.

To better identify the differential mortgage lending patterns among banks that serve the same local neighborhoods, in our second regression framework we limit the study to census tracts designated as a bank’s assessment areas under CRA regulation (12 CFR § 25.41). A bank’s lending activity within its assessment areas is the most critical factor in CRA evaluation (Avery et al. 2003; Saadi 2020). Census tracts are small county subdivisions with homogeneous demographic attributes like household income (McKinnish et al. 2010), so mortgage demand within a tract is likely to be more homogeneous. We examine the effect of ESG on banks’ mortgage lending share in a given CRA-assessment tract each year. We estimate a within-tract-year regression:

$$ MGISSUANCE\_{T}_{ity}\left( MGNUMSHR\_{T}_{ity}\right)={\beta}_1 ES{G}_{iy}+{\beta}_2 ES{G}_{iy}\times TRACTPOVERT{Y}_{ty}+{\beta}_3 DEPCNTYSH{R}_{icy}+{\chi}_{iy}+{\gamma}_{ty}+{\lambda}_i+{\epsilon}_{ity} $$

(2)

where subscript t represents a census tract. The unit of observation is the bank-tract-year, and γ_ty represents tract×year fixed effects. Identification is based on the difference in mortgage issuance among banks with different ESG scores within a given tract-year. We use two dependent variables: MGISSUANCE_T is an indicator variable equal to one if a bank extends at least one home-purchase mortgage loan in a tract-year, and zero otherwise; and MGNUMSHR_T is the number of a bank’s home-purchase mortgage loans in a tract-year as a percentage of the total number of home-purchase mortgage loans in that tract-year. We find similar results when using a bank’s mortgage share in the tract-year measured in total loan dollars, which we do not report for brevity. All other regression variables are defined as in eq. (1), with the only difference being that we measure the poverty rate at the tract-year level, denoted by TRACTPOVERTY. As in eq. (1), the coefficient of interest is β₂ on the interaction term ESG×TRACTPOVERTY. A negative (positive) β₂ indicates that high-ESG banks lend fewer mortgages in poor tracts than low-ESG banks. Standard errors are double clustered at the tract and bank level.

3.3 Summary statistics

Table 1, Panel B reports summary statistics for the bank-county-year sample. The average (median) bank-year controversy-adjusted ESG score is 0.371 (0.357) with a standard deviation of 0.108. The Pearson correlation between the raw ESG score and bank size is 84% but drops to 26% between adjusted ESG score and bank size.¹⁴

The mean (median) BANKSIZE is 16.977 (16.680), which translates to a mean (median) of USD 23.6 (17.5) billion in total assets. The average bank-year has a nonperforming loan-to-loan ratio of 1.4%, a Tier1 risk-based capital ratio of 12.3%, a deposit-to-loan ratio of 1.07, a large time deposits-to-deposits ratio of 8%, and a commercial loans-to-loans ratio of 57.3%. The average trailing two-year average loan growth is 12.8%, suggesting that the sample bank-years are in a credit expansion phase. About 8% of the sample bank-years have marketing expenses below the reporting threshold—only marketing expenses that exceed both $100,000 and 7% of the “other non-interest expenses” category on FR Y-9C are reported—and we set MARKETING for these banks to zero. Among bank-years with marketing expenses, the mean (median) value of marketing expenses-to-total non-interest expenses is 2.6 (2.4) percent.¹⁵

The mean (median) values of MGNUMSHR and MGAMTSHR are 0.034 (0.014) and 0.035 (0.014), respectively. Banks issue on average about 3.4 (3.5) percent of the total number (amount) of home mortgage loans made in that county-year. The median value of DEPCNTYSHR is zero, indicating that banks do not collect any deposits but issue home mortgages in more than half of the county-years. The mean (median) poverty rate of a county-year is 15.0 (14.4) percent.

Panel C reports summary statistics for the bank-tract-year analysis. The mean (median) value of MGISSUANCE_T is 0.286 (0). The zero median indicates that banks make no home mortgage loans in more than half of their CRA assessment tracts in a given year; the mean indicates that the banks issue a conventional home-purchase mortgage loan in 28.6% of the tract-years. The mean (median) value of MGNUMSHR_T is 0.025 (0); i.e., banks on average originate 2.5% of their assessment tract’s total home mortgage loans. The mean (median) poverty rate of a census tract is 14.1 (10.1) percent. About 1.3% and 0.2% of the observations are in middle-income rural tracts classified as distressed and underserved under the CRA, respectively.

We inspect the behavior of bank- and county-year variables across ESG ratings. Each year we sort banks into quintiles by ESG ratings, with the bottom (top) quintile comprising banks with the lowest (highest) ESG ratings. We then pool each annual quintile across years, forming five aggregate ESG-ranked buckets, and plot the mean and median values of select variables within each bucket in Fig. 1. The upper left graph in Panel A shows a monotonic increase in bank ESG ratings going from the bottom ESG quintile to the top ESG quintile, validating our sorting procedure. The upper right graph in Panel A reveals a positive relation between ESG and bank size in the bottom three ESG quintiles but a negative relation between the two variables in the top two quintiles. The downward slope in the top two quintiles is driven by the larger banks’ drawing more public scrutiny, reporting more controversies, and thus having their raw ESG score lowered more by Refinitiv. In the bottom left figure in Panel A, there is no clear relation between ESG and the percentage of nonperforming loans, which suggests that loan portfolio quality is similar for high- and low-ESG banks. The bottom right panel displays a noticeable downward trend in the Tier1 risk-based capital ratio going from the bottom to the top ESG quintile.

Fig. 1

Bank and county characteristics across bank ESG quintiles. The figures plot the mean and median values of select variables for each quintile rank formed based on banks’ ESG scores from Refinitiv

Turning to county-year characteristics in Panel B, ESG exhibits a largely negative relation with county poverty rate, which suggests that high-ESG banks lend more in rich counties, before controlling for confounding factors. ESG is positively correlated with banks’ mortgage lending share in both number and dollar amount and with banks’ deposit shares, but these relations all turn negative between the fourth and fifth ESG quintiles.

4 Main results

4.1 Visual evidence

Figure 2 shows maps of U.S. counties based on the counties’ average poverty rate (upper panel) and the weighted-average ESG ratings of banks issuing home-purchase mortgage loans in the counties (lower panel). We use the bank’s mortgage lending amount share of a county’s total mortgage lending amount as the weight when constructing the ESG map. In the poverty (ESG) map, redder counties have lower poverty rates (higher ESG scores). There is a strong correlation in redness between the two graphs; i.e., counties with redder shades in the poverty map (rich counties) are redder in the ESG map (counties populated by high-ESG banks). Banks with high ESG ratings tend to finance fewer mortgages in poor counties than banks with low ESG ratings.

Fig. 2

County poverty and county average bank ESG. Panel A provides a map of average poverty rates of all U.S. counties between 2015 and 2018, with redder shades indicating counties with lower poverty rates. Panel B provides a map of the weighted average ESG scores of banks issuing home-purchase mortgage loans in the counties, with banks’ mortgage lending amount share in the county as the weight

4.2 Bank-county-year level regression results

In Table 2, we examine the within-county-year variation in mortgage lending among banks with differing ESG ratings using eq. (1). All four columns include county × year fixed effects, with the last two columns also including bank fixed effects. Panel A (B) uses MGNUMSHR (MGAMTSHR) as the dependent variable. In Panel A, columns (1) and (3), the coefficient on ESG is statistically and economically insignificant, suggesting that, on average, a bank’s ESG has no significant relation with its mortgage lending share in a county. Adding bank fixed effects increases the adjusted R² of the regression by more than 10%, from 0.462 to 0.518, underscoring the importance of controlling for unobserved persistent differences across banks.

Table 2 Bank ESG and home-purchase mortgage origination—bank-county-level analysis

In columns (2) and (4), the coefficients on the standalone ESG are positive, suggesting that when a county’s poverty rate is zero, bank ESG has a positive influence on home mortgage lending share. The coefficient on ESG × CNTYPOVERTY is negative in both columns, suggesting that the positive effect of ESG at zero county-poverty rate weakens as the poverty rate goes up. To put the economic magnitude of ESG’s effect into perspective, we consider the coefficient on ESG × CNTYPOVERTY in column (4), which equals −0.323. This estimate indicates that in a county-year with a poverty rate at the 25th percentile of the distribution (0.109), an interquartile increase in a bank’s ESG rating (0.118) is associated with a 12-basis-point (bps) increase in mortgage lending share (=[0.046 × 0.118] + [−0.323 × 0.118 × 0.109]); but, in a county-year with a poverty rate at the 75th percentile of the distribution (0.183), an interquartile increase in a bank’s ESG score is associated with a 15-bps reduction in mortgage lending share (=[0.046 × 0.118] + [−0.323 × 0.118 × 0.183]), which represents a 4.4% drop relative to the sample mean of MGNUMSHR (0.034).¹⁶

Panel B, which uses MGAMTSHR as the dependent variable, has results similar to Panel A. According to the estimates in columns (1) and (3), bank ESG has no discernible effect on mortgage lending share. Yet, there is much variation in the ESG effect conditioned on a county’s poverty rate, as reflected in the negative and significant coefficients on ESG × CNTYPOVERTY in columns (2) and (4). The coefficient on ESG × CNTYPOVERTY in column (4) is equal to −0.356, which indicates that in a county-year with a 25th percentile poverty rate of 0.109, an interquartile increase in a bank’s ESG rating is associated with a 16-basis-point increase in mortgage lending share (=[0.053 × 0.118] + [−0.356 × 0.118 × 0.109]); in a county-year with a 75th percentile poverty rate of 0.183, an interquartile increase in a bank’s ESG score is associated with a 14-basis-point reduction in mortgage lending share (=[0.053 × 0.118] + [−0.356 × 0.118 × 0.183]), which is a 4.1% drop relative to the sample mean of MGAMTSHR (0.035).

Figure 3 plots the effect of ESG on home mortgage lending share across decile ranks of the county poverty rate. To construct this figure, we estimate a modified version of eq. (1) replacing the continuous county-poverty-rate variable with dummies indicating each of 10 decile ranks formed annually. The combined coefficients of ESG and each ESG × Poverty Decile Dummy are plotted—for example, the ESG effect in the bottom poverty decile is the standalone coefficient on ESG (with the poverty decile dummy numbered 0). The effect of ESG on home mortgage lending share, in terms of both quantity and dollar amount, is almost monotonically decreasing in the county poverty rate. Starting from the fifth poverty quintile (as displayed by the dashed green line), ESG has a negative effect on mortgage lending share—high-ESG banks lend less than low-ESG banks in poorer counties, all else equal.

Fig. 3

Estimated ESG effect on mortgage lending share by county poverty decile. This figure reports the effect of ESG on banks’ home mortgage lending share by county poverty decile. We estimate a modified version of eq. (1), replacing the continuous county-poverty-rate variable with dummies indicating each of the 10 decile ranks formed annually by county poverty rate. The combined coefficients of ESG and each ESG × Poverty Decile Dummy are plotted—the ESG effect in the bottom poverty decile is the standalone coefficient on ESG when the poverty decile dummy is numbered zero

In Online Appendix Table A3, we rerun the analyses dropping the “Big 4” banks—Citigroup, JPMorgan Chase, Bank of America, and Wells Fargo—or multinational banks. The results stand, confirming that bank size and complexity do not drive the observed ESG–poverty lending disparity. In Table A4, we show that high-ESG banks do not have lower deposit market shares in low-income ZIP codes within the same county and are no more likely to shutter branch offices in poor counties than low-ESG banks. Therefore, our results are not driven by banks’ differential deposit networks. In Table A5, we link bank ESG ratings to small business lending and, consistent with our main results, find that high-ESG banks lend less to small businesses in poor localities than low-ESG banks. In light of the wide divergence in ESG ratings from different data vendors (e.g., Christensen et al. 2022), in Table A7 we re-estimate our main regressions using alternative ESG data from Bloomberg, S&P Global, and MSCI/KLD. We observe the same ESG–poverty lending disparity for all three alternative ESG data sources.

4.3 Bank-CRA assessment tract-year-level regression results

High-ESG banks might not be located or have branches in poor localities. For example, a bank headquartered and branched on the West Coast may have low mortgage exposure in low-income regions on the East Coast due to constraints from geographical diversification, information frictions about borrowers, and competition in those mortgage markets. To isolate banks’ local mortgage lending, we further examine the lending patterns in banks’ CRA assessment tracts. The CRA requires banks to delineate census tract areas wherein the federal regulators will evaluate the banks’ record of helping the credit needs of local communities (12 CFR 228.41). The assessment areas typically cover regions in which banks hold deposit-gathering facilities (e.g., branches, ATMs) and the surrounding geographical areas.¹⁷ We identify banks’ CRA assessment tracts from the FFIEC CRA Disclosure File and estimate eq. (2).

Fig. 4 shows maps of the poverty rate and bank ESG for census tracts in two counties—Bucks County in Pennsylvania (Panel A, near Philadelphia) and Hudson County in New Jersey (Panel B, near New York City). We compute the average ESG scores of banks issuing mortgages in the census tract. Unlike in Fig. 2, because few banks lend in census tracts in a given tract-year, we do not use banks’ mortgage lending amount share as the weight. Tracts with redder shades have lower poverty rates (left panel) and higher average bank ESG (right panel). We see that redness overlaps the poverty rate and bank ESG maps for both counties; that is, even within a county, low-ESG banks issue more home mortgage loans in poor neighborhoods than high-ESG banks. These maps illustrate a pervasive pattern: high-ESG banks do not excel at providing mortgage credit to homebuyers in low-income neighborhoods.

Fig. 4

Tract poverty and tract-lending-bank ESG – two examples. Panel A provides maps of the poverty rates (left panel) and the average ESG scores of banks issuing mortgages in the census tracts (right panel) in Bucks County, PA. Panel B provides the same set of maps for census tracts in Hudson County, N.J. Tracts with redder shades have lower poverty rates (left panel) and higher average bank ESG (right panel)

Table 3 reports the regression estimates. We begin by estimating ESG’s effect on the incidence of banks issuing at least one home mortgage loan in a tract-year using MGISSUANCE_T as the dependent variable. The regression is estimated on a sample of 2,978,042 bank-tract-years from 738,843 tract-years (averaging four banks per tract-year). In column (1), the coefficient on ESG is insignificant statistically and economically, suggesting that a bank’s ESG rating has no bearing on whether the bank issues mortgage loans in a census tract. In column (2), the coefficient on ESG × TRACTPOVERTY equals −0.751 and is statistically significant (p-value = 0.004), indicating that high-ESG banks issue fewer home-purchase mortgage loans in poor tracts than their low-ESG counterparts that serve the same tracts. In tracts with a 25th percentile poverty rate of 0.051, an interquartile increase in ESG (0.117) is associated with a 1.4% increase (=[0.160 × 0.117] + [−0.751 × 0.117 × 0.051]) in banks’ propensity to lend. However, in tracts with a 75th percentile poverty rate of 0.193, the same 0.117 interquartile increase in ESG is associated with a 17.6-bps reduction (=[0.160 × 0.117] + [−0.751 × 0.117 × 0.193]) in banks’ propensity to lend.

Table 3 Bank ESG and home-purchase mortgage origination – bank-CRA assessment tract-level analysis

In columns (2) and (3), we interact ESG with an indicator reflecting whether a census tract is in distress (DISTRESSED) or is underserved (UNDERSERVED), as classified by federal banking agencies under the CRA. The coefficients on ESG × DISTRESSED and ESG × UNDERSERVED are both negative and statistically significant, suggesting that high-ESG banks are less likely than low-ESG banks to extend home mortgage loans in distressed and underserved communities.

In Panel B, we test the effect of ESG on the market share of banks’ home mortgage lending in assessment tracts. We restrict the sample to banks that issue at least one mortgage loan in the tract-year. The regression sample comprises 601,603 bank-tract-years from 243,196 tract-years. The coefficient on ESG in column (1) is reliably negative (coefficient = −0.007; p-value = 0.048), suggesting that high-ESG banks, on average, hold a lower proportion of a tract’s home mortgage lending volume than do low-ESG banks. This average effect, as revealed in column (2), derives mainly from poor communities. Specifically, the coefficient on standalone ESG is statistically insignificant, while the coefficient on the interaction term ESG × TRACTPOVERTY is negative and statistically significant (coefficient = −0.086; p-value = 0.013). This suggests that high-ESG banks lend a smaller portion of a poor tract’s total mortgage loans than low-ESG banks. In column (3), the coefficient on ESG × DISTRESSED is negative and significant (coefficient = −0.046; p-value = 0.002), suggesting that high-ESG banks lend less than low-ESG banks in distressed areas. In column (4), the coefficient on ESG × UNDERSERVED is statistically insignificant (unlike that in Panel A, column (4)), suggesting that the negative effect of distressed rural communities resides mainly in the incidence of loan issuance.

Figure 5 plots the estimated effect of ESG on banks’ propensity to extend a home mortgage loan across census tracts ranked by poverty decile. The figure reveals a monotonic decrease in high-ESG banks’ home mortgage origination likelihood relative to that of low-ESG banks moving from the richest to the poorest tracts. In counties with poverty rates beyond the fourth decile of the poverty distribution, high-ESG banks are less likely to lend than low-ESG banks.¹⁸

Fig. 5

Estimated ESG effect on mortgage loan issuance by CRA assessment tract poverty decile. This figure presents the effect of bank ESG on home mortgage lending by decile ranks of tract poverty rates. We first estimate a modified version of eq. (2), replacing the continuous tract poverty rate variable with dummies for each of the decile groups formed annually by tract poverty. We then plot the combined coefficients of ESG and the two-way interaction of ESG and each of the decile dummies, with the estimated ESG effect for the bottom decile being the coefficient on the standalone ESG

5 Additional analyses

5.1 Hurricanes, social wash, and mortgage lending

We use severe hurricanes as a quasi-exogenous setting where access to mortgage credit is particularly important for households (as they seek to recover from disaster losses) but banks might be less willing to lend (because of increased financial risk). These events let us examine whether banks engage in social wash when social responsibility matters the most. After a severe hurricane, families need to repair or replace destroyed properties. In poor communities, however, many properties are either uncovered or inadequately covered by homeowners’ and flood insurance, leaving homeowners little choice but to borrow. The extra borrowing is likely to put poor households further in debt, and it increases their mortgage default risk, which decreases lenders’ willingness to lend. Moreover, perceived hurricane risk can increase after a realized event, further eroding banks’ credit supply (Bin and Polasky 2004; Morse 2011). On the other hand, banks have increased incentives to show they are socially conscious after natural disasters (like Hurricane Katrina) because these events draw a lot of publicity and media attention. To the extent that banks view the benefits of ESG promotion as outweighing the small costs of implementing symbolic actions and being disciplined for misleading ESG disclosures, social wash will arise.

We use an event-study framework, with the event window stretching from three years before to three years after a severe hurricane hits a given county. We restrict the sample to banks that issued mortgage loans in the county before the hurricane and ask whether high-ESG banks are more (or less) likely than low-ESG banks to cut lending after the hurricane. We collect severe hurricane events from the Costliest U.S. Tropical Cyclones list provided by the National Oceanic and Atmospheric Administration (NOAA). We identify counties affected by the hurricanes from the Federal Emergency Management Agency (FEMA) Disaster Declaration Summaries file.

We create three indicator variables denoting each of the three years after a hurricane—Year +1, Year +2, and Year +3—and modify eq. (1) by interacting the three dummies with ESG, CNTYPOVERTY, and ESG × CNTYPOVERTY. In Table 4, column (1), we focus on banks’ decision to issue a mortgage loan: the dependent variable MGISSUANCE is an indicator set to one if the bank issues a mortgage loan in the county-year. Each of the triple-interaction terms compares the propensity of high-ESG banks to lend in poorer counties in a post-hurricane year to that of low-ESG banks. The estimates show that high-ESG banks are more likely than low-ESG banks to stop lending in poor neighborhoods hit by severe hurricanes. This effect exists throughout the post-event period (though the statistical significance of the triple interaction term for year 3 is weaker), with the high-ESG banks’ retreat peaking in year 2. Column (3) repeats the test excluding Hurricane Katrina (the most damaging hurricane), and the results continue to hold.

Table 4 Bank ESG, home mortgage lending, and severe hurricanes

In columns (2) and (4), we limit the sample to banks that issued at least one mortgage loan both before and after a hurricane. We ask whether, among banks that continue to lend in affected areas after a hurricane, high-ESG banks reduce their mortgage lending share in poorer areas more than low-ESG banks. The dependent variable here is MGAMTSHR. As reported, high-ESG banks appear to cut their mortgage lending share in poor areas the most in year 2, although this effect is not statistically significant at conventional levels. The collective evidence is consistent with the social wash effect: banks with higher ESG ratings are quicker to withdraw from poor areas after a severe hurricane, when mortgage access is often crucial to families’ recovery.

5.2 Instrumental variable approach

To assess the causal impact of ESG disclosures, we use an instrumental variable (IV) that creates plausibly exogenous variation in banks’ ESG ratings. Firms located in Democrat-leaning areas are more likely to espouse ESG policies than firms located in Republican-leaning areas (Di Giuli and Kostovetsky 2014). We instrument a bank’s ESG rating in a given year with the percentage of voters in the bank’s headquarter state that voted for the Democratic presidential candidate in the last presidential election, which we call DEMPCT. Because a state’s party alignment shifts over time, the instrument induces time-series variation in banks’ ESG alignment. Moreover, banks’ headquarter locations are predetermined and unlikely to change because of an election outcome, bolstering the exogenous nature of the instrument. We obtain states’ presidential voting data from Dave Leip’s Atlas of U.S. Presidential Elections. A critical assumption of this IV approach is that states’ presidential voting pattern affects banks’ mortgage lending behavior only through its effect on banks’ ESG policies and not through other mechanisms. While this exclusion restriction is not testable, we know of no prior evidence that borrowers’ political ideology can directly shape banks’ mortgage lending decisions. In Online Appendix Table A8, we rule out a potential violation of the exclusion restriction: banks headquartered in Republican-leaning states could disproportionately service poorer rural counties, and people wishing to buy properties there may be more inclined to deal with banks that share similar political preferences.¹⁹

We replicate the bank-county-year analysis in Table 2 using the 2SLS model. Because the outcome variables (MGNUMSHR and MGAMTSHR) are at the bank-county-year level while ESG is a bank-year measure, the first- and second-stage models have a different unit of observation. In the first stage, we model ESG as a function of DEMPCT along with bank-year-level controls and year fixed effects. In the second stage, we regress the outcome variables on the fitted values from the first-stage models, the interaction of the fitted values and CNTYPOVERTY alongside control variables, bank fixed effects, and county × year fixed effects.²⁰

Table 5 reports the 2SLS results. Column (1) reports the first-stage regression result. The coefficient on DEMPCT confirms that banks in states that recently voted for the Democratic presidential nominee have higher ESG ratings than banks in states that voted for the Republican nominee. The first-stage F-statistic is 22, greater than the 10 that Staiger and Stock (1997) suggest researchers use as a threshold to identify strong instruments. Columns (2) through (5) report the second stage estimates. In columns (2) and (4), we do not interact ESG with CNTYPOVERTY. Unlike in the OLS regressions, the coefficient on standalone ESG has a negative sign, though it is not statistically significant at conventional levels. In columns (3) and (5), we add the interaction term ESG × CNTYPOVERTY, and the coefficients are negative and statistically significant in both regressions. This result confirms that increases in banks’ ESG disclosure intensity correspond to banks reducing their share of mortgage lending in low-income neighborhoods. In an untabulated analysis, we replicate the tract-year-level analysis in Table 3 using this 2SLS framework, and our inferences remain the same.

Table 5 2SLS instrumental variable approach

5.3 Does CRA enforcement undo banks’ social wash behavior?

We examine the efficacy of CRA enforcement in mitigating social wash. The CRA requires federal bank regulators to assess commercial banks’ performance in meeting local communities’ credit needs, especially in poor neighborhoods. Banks are assigned one of four statutory ratings upon the completion of a CRA exam: “Outstanding,” “Satisfactory,” “Needs to Improve,” or “Substantial Noncompliance” (12 U.S.C. 2906(b)(2)). The first two ratings are considered passing, and the last two are failing, although banks rarely receive failing grades. The Financial Institutions Reform, Recovery, and Enforcement Act of 1989 required regulators to publicly disclose banks’ CRA ratings. We retrieve depositor institutions’ CRA ratings from the FFIEC Interagency CRA Rating File. We create an indicator variable OUTSTANDING, which equals one for banks that received an “Outstanding” rating in their most recent CRA exams. For bank holding companies with multiple CRA depository institutions, all subsidiary institutions must receive an “Outstanding” rating for the holding company to have OUTSTANDING coded one. About 20% of bank-years in the sample have such a rating. We augment eqs. (1) and (2) by interacting ESG and CNTYPOVERTY (TRACTPOVERTY) with the OUTSTANDING dummy.

By monitoring and grading bank lending records in assessment tracts, CRA examiners could narrow the ESG disclosure–poverty lending gap. Banks that fail CRA exams are barred from mergers and acquisitions, new branch openings, and other expansions until their record improves. If CRA enforcement is robust, we expect ESG ratings and mortgage lending to be better aligned among banks that do well in the CRA examination. However, skeptics argue that CRA ratings are inflated (very few banks fail), subject to regulatory capture, and subjective, and that banks that game the ESG scoring system can easily gain CRA credit. Under this view, the disconnect between bank ESG and mortgage lending in low-income areas will not vary with CRA ratings.

Table 6 reports the regression results. Columns (1) and (2) display the results for the bank-county-year data, and column (3) the results for bank-tract-year data. The coefficients on ESG × POVERTY are negative and statistically significant in all three columns, suggesting that among banks without an Outstanding CRA rating, high-ESG banks finance fewer home mortgages than low-ESG banks in poor neighborhoods. This disconnect is mitigated among banks with an Outstanding CRA rating. The coefficient on the triple interaction term ESG × POVERTY × OUTSTANDING is positive across all three regressions, with p-values ranging from 0.096 (column 2) to 0.176 (column 1). Receiving an “Outstanding” CRA rating offsets high-ESG banks’ lending contraction in poor areas by about 50%. Thus, the CRA examination and enforcement blunts, but does not eliminate, banks’ social wash behavior.

Table 6 Is the social wash effect mitigated by CRA enforcement?

In Online Appendix Table A9, we compare high-ESG banks’ mortgage lending behavior with that of low-ESG banks after CRA rating downgrades; i.e., when the bank’s CRA rating drops, for example, from “Satisfactory” to “Needs to Improve.” We find that high-ESG banks increase their mortgage lending in poor neighborhoods more vigorously after CRA downgrades than their low-ESG peers. Our interpretation is that when CRA downgrades expose high-ESG banks’ lackluster community lending (which is not properly conveyed in unaudited ESG disclosures), the banks quickly channel credit into low-income areas to recoup their lost social credentials.

5.4 Mortgage loan application analysis

We probe the relation between banks’ ESG ratings and their lending decisions at the individual loan application level. We ask whether high-ESG banks are more likely than low-ESG banks to reject a mortgage loan application and charge higher prices on accepted mortgages across U.S. counties. The HMDA provides applicant data such as income, ethnicity, race, and loan data such as the mortgage amount. These attributes are correlated with mortgage decisions, and by controlling for them in the regressions, we seek to disentangle the incremental effect of ESG from application-specific attributes. We run the following loan-application-level regression:

$$ {Y}_{mi}={\alpha}_1 ES{G}_i+{\alpha}_2 ES{G}_{iy}\times POVERTY+\beta {\sigma}_m+{\chi}_{iy}+{\alpha}_{cy}+{\lambda}_i+{\epsilon}_{mi} $$

(3)

where subscripts m, i, c, and y represent mortgage applicant, bank, county, and year, respectively. The dependent variable is either an indicator variable reflecting whether a loan is denied by the bank (DENIAL) or an indicator variable for higher-priced loans (HIGHPRICE) whose annual percentage rates exceed the thresholds established by the HMDA, also called “subprime loans.” The control variables include the borrower’s debt-to-income ratio (DTI), the natural log of income in thousands of dollars (INCOME), whether a loan exceeds the conforming loan size limit and thus cannot be sold to government-sponsored entities (JUMBO), whether a loan is a first lien (FIRSTLIEN), and whether the applicant is Hispanic (HISPANIC), African American (BLACK), or female (FEMALE). As in eq. (1), we control for bank characteristics, bank fixed effects, and county × year fixed effects.

Table 7 reports the results. In columns (1) and (2), we examine the effect of ESG on banks’ propensity to deny a loan, with and without the interaction term ESG × CNTYPOVERTY. The coefficient estimates indicate that after controlling for borrower and loan attributes, the propensity to deny a mortgage application increases with the county’s poverty rate faster for high-ESG banks than for low-ESG banks.

Table 7 Individual mortgage application decisions

In columns (3) and (4), we examine the effect of ESG on banks’ issuance of high-priced loans. The coefficients on ESG and ESG × POVERTY are statistically insignificant in both columns, suggesting that after controlling for borrower and loan attributes, high-ESG banks charge similar interest rates on mortgage loans as low-ESG banks and this (lack of) difference does not vary with the poverty rate of the county where the property is located. This non-result suggests that unobserved credit risk metrics like credit scores are unlikely to drive the earlier loan denial results: if high-ESG banks’ higher denial rates in low-income areas are driven by less creditworthy borrowers, then these same credit risks should be reflected in higher interest rates, which is not the case. From this loan-application-level analysis, we can infer that the disconnect between banks’ ESG disclosures and mortgage lending derives mainly from the banks’ differential loan acceptance/denial decisions, not from loan pricing.²¹

5.5 Using the “S” component of ESG

Banks’ mortgage lending in low-income areas contributes mainly to the social aspect of bank ESG. To sharpen our inferences, we isolate the “S” part of ESG and redo the main regressions replacing ESG with S. Table 8 columns (1) and (2) report the bank-county-level regressions, and column (3) reports the bank-tract-level regressions. We consistently find that banks with high social ratings cut lending more than banks with low social ratings in poor communities.

Table 8 Using “S” of “ESG”

5.6 Evaluating alternative explanations

5.6.1 Do high-ESG banks make better-quality loans than low-ESG banks?

An alternative explanation is that by not lending in poor areas, high-ESG banks reduce loans to people who cannot afford them, reducing future loan defaults. Since mortgage defaults have negative social and economic consequences, it can be argued that high-ESG banks’ socially desirable prudent underwriting standards create an appearance of credit rationing in poor areas.

We test whether high-ESG banks make better-quality mortgage loans than low-ESG banks. If the alternative explanation is correct, we would expect high-ESG banks to have fewer mortgage payment defaults than low-ESG banks. We capture banks’ mortgage lending quality using both the amount of home mortgage loans classified as non-performing loans (NPLMG) and the amount of home mortgage loans that were charged-off net of recoveries (NCOMG) in a given year, normalized by the total amount of loans at the beginning of the year. NPLMG is a stock measure of loan portfolio quality, reflecting loans that are 90 days past due or that are placed on nonaccrual status. NCOMG is a flow measure of loan portfolio quality, reflecting confirmed mortgage losses in a given year that banks charge off from their balance sheets. We conduct the following regression at the bank-year level:

$$ Mortgage\ Qualit{y}_{iy+1}={\beta}_1 ES{G}_{iy}+ Contro{l}_{iy}+ Bank\ FE+{\epsilon}_{iy}, $$

(4)

where the dependent variable is either NPLMG or NCOMG. We include the same set of bank-year control variables as in the main regression, except that we drop current-period NPL from the model since the dependent variable is next-period mortgage NPL (though our results do not change much if we retain current-period NPL). We include bank fixed effects to control for time-invariant unobserved bank characteristics. The coefficient of interest is β₁, which indicates the effect of a bank’s current-year ESG on the bank’s delinquent mortgage loans in the following year.²²

Table 9 reports the results. For both dependent variables, the coefficient on ESG is small and statistically insignificant, suggesting that high-ESG banks do not experience more or fewer mortgage delinquencies and charge-offs than low-ESG banks. We infer that there is no difference in mortgage lending standards among banks with differing ESG ratings and that the observed ESG–mortgage lending disparity is not explained by differential mortgage lending standards.

Table 9 Do high-ESG banks make better mortgage loans than low-ESG banks?

5.6.2 Do ESG ratings respond to banks’ community lending?

One may suspect that, because ESG ratings reflect a broad spectrum of corporate actions (from environmental efforts to employee treatment to community engagement), a bank’s low-income lending constitutes only a tiny fraction of its ESG rating. So, even if a bank increased its low-income lending, these efforts may not show up in its ESG ratings, in which case our use of noisy ratings would induce bias. We believe that ESG ratings groups are likely to scrutinize a firm’s main business activities for ESG efforts. Just as oil-and-gas firms can directly help the environment by cutting their carbon dioxide emissions, banks can fulfill their social obligations by making one of their main products, mortgage loans, more accessible to low-income households.

We test whether Refinitiv ESG ratings incorporate banks’ low-income lending. We regress a bank’s annual ESG rating on a summary measure of the bank’s low-income lending intensity in that year, LLI (LLI is the weighted average of the bank’s mortgage lending share across counties weighted by county poverty rates); the bank’s CRA exam rating; the interaction of the two measures; and control variables as well as bank and year fixed effects. The bank fixed effects help identify how a bank’s ESG rating changes in response to changes in its low-income lending record.

In Table 10 columns (1) and (2), we include individually LLI and CRA, and in column (3) we add their interaction. As the first two columns report, a bank’s low-income lending intensity and CRA ratings are independently and positively associated with its ESG ratings, suggesting that Refinitiv ESG ratings reflect both lending metrics. In column (3), the coefficient on LLI is negative, the coefficient on CRA is insignificant, and the coefficient on CRA × LLI is positive, suggesting that the impacts of low-income lending intensity and CRA on bank ESG rating feed off each other. Put another way, ESG ratings incorporate changes in LLI and CRA ratings more forcefully when the two metrics are consistent with each other.

Table 10 Do ESG ratings respond to CRA ratings and low-income mortgage lending?

5.6.3 Do high-ESG banks substitute public welfare investment for mortgage loans in poor areas?

Outside mortgage loans, banks can make community development investments, such as Low-Income Housing Tax Credit (LIHTC) or other affordable housing projects, to revitalize their communities. So, another alternative explanation for our results is that high-ESG banks substitute community investments for mortgage lending in low-income neighborhoods. Using data on banks’ public welfare investments (PWIs) authorized by law (12 USC §24), we show in Online Appendix Table A10 that high-ESG banks cut mortgage lending more in poor counties when they have fewer public welfare investments there. This result indicates a complementarity of banks’ mortgage loans and public welfare investments in poor localities, which is inconsistent with the alternative view.

6 Conclusion

We find that high-ESG banks issue fewer home mortgages, in both number and dollar amount, in poor counties than do low-ESG banks. Even within the same county, high-ESG banks lend less in poor census tracts compared to low-ESG banks. Our results are not driven by time-varying credit demand factors or by persistent differences across banks and are robust to instrumenting banks’ ESG disclosures with headquarter states’ political orientation and to using ESG ratings from multiple data providers. The evidence hints at social wash: banks project an image of social altruism while shunning tangible actions that create true social good.

We uncover additional patterns concerning the social wash effect. First, we find that, in poor areas hit by big hurricanes, high-ESG banks are more likely to halt lending than low-ESG banks, despite mortgage credit being critical to families’ recovery in the storms’ aftermath. Second, we find that CRA enforcement is somewhat effective in undoing social wash, as high-ESG banks with an “Outstanding” performance rating shrink lending less in poor areas. We also report that high-ESG banks are quicker to increase low-income lending after a CRA downgrade than low-ESG banks.

An alternative view is that high-ESG banks have better lending standards, so their appearance of credit rationing in poor areas is attributable to denial of borrowers who are not creditworthy. Yet our tests reveal little difference in the fraction of banks’ mortgage loans that are delinquent or charged-off between high- and low-ESG banks, suggesting that high-ESG banks do not have better lending standards than low-ESG banks. Another criticism is that banks’ home mortgage lending constitutes only a small fraction of their overall ESG endeavors, so ESG ratings may not change even when banks lend more in low-income areas. We show that changes in banks’ low-income lending intensity, along with changes in CRA ratings, predict changes in ESG ratings. Our evidence is also inconsistent with high-ESG banks substituting public welfare investments (like affordable housing projects) for mortgage origination in low-income areas. Our skepticism about high-ESG rating banks’ actual performance is shared by mutual funds that have recently gone public about shunning green bonds issued by prominent banks and bolstered by reports of widely different ESG scores assigned by different raters to the same bank (e.g., Wirz 2021; Eaglesham 2022).

Appendix

Table 11 Variable Definitions