A. Data

To construct our sample, we start with all corporate syndicated loans issued to the U.S. firms that have financial information available from Compustat during the period 1987–2017. Data on corporate loans are collected from the DealScan database maintained by the Loan Pricing Corporation, which contains information about loan pricing and various loan terms at origination, such as loan size, use of collateral, and covenants. We match the loan data with Compustat firms using the most updated DealScan–Compustat link file, maintained by Michael Roberts and Wharton Research Data Services (Chava and Roberts (Reference Chava and Roberts2008)). Loans with information missing on key terms, such as loan spread, lender identity, or loan amount, are excluded from the sample. We also remove loans issued to financial firms (SIC 6000–6999) and regulated utilities (SIC 4900–4949).

We obtain data on major natural disasters from the Spatial Hazard Events and Losses Database for the United States (SHELDUS) at Arizona State University. For each event, SHELDUS provides information on the event dates, disaster types, property damages, fatalities, and affected county locations of major natural disasters in the United States. To ensure that an event is salient enough, we focus on disasters with total estimated damages above $1 billion 2017 dollars and with a duration less than 30 days (Barrot and Sauvagnat (Reference Barrot and Sauvagnat2016)). This filtering procedure leaves us with 40 major disasters during 1987–2017. Table 1 reports basic information about these disasters, such as disaster name, disaster date, and the number of counties affected by each disaster and counties neighboring to them. The events are not clustered but rather dispersed over our sample period.

TABLE 1Disaster Basics

TABLE 1 Long description

The table contains four columns: Disaster, Date, Number of Affected Counties, and Number of Neighboring Counties.

Key entries include:

* Hurricane Hugo, September 1989: 190 affected, 186 neighboring.

* Earthquake Loma Prieta, October 1989: 8 affected, 12 neighboring.

* Hurricane Bob, August 1991: 85 affected, 61 neighboring.

* Wildfires Oakland Hills, October 1991: 1 affected, 5 neighboring.

* Hurricane Andrew, August 1992: 78 affected, 118 neighboring.

* Hurricane Iniki, September 1992: 1 affected, 5 neighboring.

* Blizzard Storm of the Century, March 1993: 938 affected, 180 neighboring.

* Earthquake Northridge, January 1994: 1 affected, 5 neighboring.

* Tropical Storm Alberto 1994, July 1994: 89 affected, 95 neighboring.

* Hurricane Opal, October 1995: 337 affected, 331 neighboring.

* Blizzard/Flooding, January 1996: 587 affected, 288 neighboring.

* Hurricane Fran, September 1996: 207 affected, 173 neighboring.

* Hurricane Georges, January 1998: 126 affected, 125 neighboring.

* Ice Storm Northeast, January 1998: 42 affected, 23 neighboring.

* Hurricane Bonnie, August 1998: 40 affected, 29 neighboring.

* Hurricane Floyd, September 1999: 300 affected, 157 neighboring.

* Tropical Storm Allison, June 2001: 201 affected, 380 neighboring.

* Wildfires Southern California, June 2003: 2 affected, 9 neighboring.

* Hurricane Isabel, September 2003: 221 affected, 135 neighboring.

* Hurricane Charley, August 2004: 82 affected, 99 neighboring.

* Hurricane Frances, September 2004: 281 affected, 269 neighboring.

* Hurricane Jeanne, September 2004: 158 affected, 146 neighboring.

* Hurricane Ivan, September 2004: 426 affected, 321 neighboring.

* Hurricane Dennis, July 2005: 267 affected, 413 neighboring.

* Hurricane Katrina, August 2005: 315 affected, 338 neighboring.

* Hurricane Rita, September 2005: 139 affected, 190 neighboring.

* Hurricane Wilma, October 2005: 24 affected, 18 neighboring.

* Flooding Midwest, April 2008: 190 affected, 294 neighboring.

* Hurricane Gustav, August 2008: 141 affected, 165 neighboring.

* Hurricane Ike, September 2008: 511 affected, 473 neighboring.

* Blizzard Groundhog Day, February 2011: 241 affected, 406 neighboring.

* Hurricane Irene, August 2011: 198 affected, 148 neighboring.

* Tropical Storm Lee, September 2011: 170 affected, 236 neighboring.

* Hurricane Isaac, August 2012: 96 affected, 74 neighboring.

* Hurricane Sandy, October 2012: 280 affected, 184 neighboring.

* Flooding/Severe Weather Illinois, April 2013: 43 affected, 108 neighboring.

* Flooding Colorado, September 2013: 8 affected, 15 neighboring.

* Tornadoes/Flooding Midwest/Southeast/Northeast, 25 April 2014: 118 affected, 275 neighboring.

* Flooding East/S C, 1 October 2015: 23 affected, 35 neighboring.

* Hurricane Matthew, 2 October 2016: 116 affected, 115 neighboring.

The final row indicates the number of unique counties: 1,913 affected and 2,209 neighboring.

We determine whether a borrowing firm is affected by a disaster event based on the actual location of its operation.Footnote ² Specifically, we obtain firms’ geographical footprints from the National Establishment Time Series (NETS) database, which provides annual establishment-level data on U.S. businesses and establishments from 1989 onward. For years prior to 1989, we use 1989 values as proxies. We match firms in our sample to NETS based on the name and address of the firm’s headquarters. Information on historical headquarters locations is obtained from Bill McDonald’s 10-X Header Database, which records firms’ reported addresses at the time of each SEC filing.Footnote ³ For firms without electronic filings prior to 1994, we assign the headquarters location based on the earliest available filing.

B. Empirical Methodology

We rely on both the occurrence of disasters and the exposure of the borrower to the disaster area to identify situations where a bank’s perception of the borrowing firm’s natural disaster risk could increase significantly. While it is highly probable that lenders’ risk perception in a disaster-affected area will change after the event, loans to firms in these areas are initially excluded from our main analysis. This exclusion is necessary as these firms’ fundamentals are likely to be affected by the disaster; including them would make it difficult to isolate the impact of salience bias on the cost of borrowing. Barrot and Sauvagnat (Reference Barrot and Sauvagnat2016) show that the effects of natural disasters on affected firms can last for more than 1 year (i.e., five quarters). Therefore, a firm in a disaster area is considered a disaster-affected firm and any loan issued to such firms within 24 months of the disaster is excluded from our analysis.

Among the remaining loans, we identify disaster-neighboring firms based on the locations of corporate establishments obtained from the NETS database. To do so, for each disaster event, we identify the neighboring counties of each disaster-affected county. Specifically, we collect information on county adjacency from the National Bureau of Economic Research and match each affected county with its five closest nonaffected counties.Footnote ⁴ Firms that derive more than 50% of their sales from these neighboring counties are categorized as disaster-neighboring firms (referred to as neighboring firms for brevity), while the rest are labeled as remote firms.

By construction, disaster-neighboring and remote firms in our sample are unlikely to be significantly and directly affected by the natural disaster. However, from the perspective of bank lenders, the perceived risk of being struck is likely to be higher for firms in the neighborhood of the disaster area. Therefore, we assess the potential impact of salience bias by comparing bank loans granted to neighboring firms following the disaster with loans to other firms.

To examine whether bank lenders respond differently to neighboring firms following a disaster, we estimate the following OLS regressions:

(1) $$ \begin{array}{l} Log{(Spread)}_{f,i,b,t}=\alpha +\beta {Neighbor}_{c,t}+{\gamma Controls}_{f,i,t-1}+{\delta}_i+{\theta}_{b,t}\\ {}+{\vartheta}_{s,t}+{\mu}_{j,t}+{\varepsilon}_{f,i,b,s,j,t.}\end{array} $$

The observation is at the loan–bank level, which means that in a loan facility with multiple lead lenders, we treat each lender as a separate observation.Footnote ⁵ This type of analysis allows us to capture heterogeneity across different bank lenders. In this specification, f denotes the loan facility, i denotes the borrowing firm, b denotes the bank, t denotes the loan initiation year, c (s) denotes the county (state) in which the borrowing firm’s headquarters is located, and j denotes the industry of the borrowing firm. The dependent variable is Log(Spread) _f,i,b,t, the natural logarithm of the all-in-drawn spread (in basis points) for loan facility f of firm i issued by bank b in year t. The explanatory variable of interest is Neighbor, an indicator variable that equals 1 if the borrowing firm is a disaster-neighboring firm within 24 months of a natural disaster, and 0 otherwise. According to Rhodes, Campbell, and Dawson (Reference Rhodes, Campbell and Dawson2004) and Murfin (Reference Murfin2012), an average loan is negotiated 2–3 months in advance of the legal effective date of the loan. To account for this time lag, we lag the loan starting date in DealScan by 3 months.Footnote ⁶

We include a number of firm characteristics that might affect loan spreads, including firm size (natural logarithm of total assets), cash holding, market-to-book ratio, financial leverage, ROA, asset tangibility, and Z-score (e.g., Lin, Ma, Malatesta, and Xuan (Reference Lin, Ma, Malatesta and Xuan2011)). All these firm characteristics are measured at the fiscal year prior to the loan initiation. To control for unobservable time-invariant heterogeneity across firms, we further include firm fixed effects (δ_i) in the regression. In addition, we control for various loan characteristics, including loan size, loan maturity, prior lending relationship, loan types (i.e., term loans, revolvers longer than 1 year, revolvers shorter than 1 year, and 364-day loans), loan purposes (i.e., general corporate purpose, refinancing, acquisition, backup line for commercial paper, and others), and a dummy variable indicating whether the loan includes a contingent performance-based pricing clause.

In our primary specification, we include bank-year fixed effects (θ_b,t) to control for time-varying heterogeneity across banks. This is important because natural disasters may affect banks’ fundamentals in different ways, such as their exposure to local deposit markets or their ability to assess borrower risk following extreme events. Bank-year fixed effects absorb these bank-level characteristics and thus control for the impact of bank fundamentals on the supply of bank loans. This approach strengthens identification by comparing lending by the same bank to firms located inside versus outside neighboring disaster-affected areas within the same year, thereby accounting for unobserved, time-varying bank characteristics that could otherwise bias the estimates. In addition, state-year fixed effects (ϑ_s,t) are included, allowing us to compare loans issued to borrowers in the same state and year. Finally, we include industry-year fixed effects (μ_j,t) to address time-varying differences across industries.

C. Summary Statistics

After requiring nonmissing values for our main control variables, we obtain a sample of 32,446 loans during the 1987–2017 period issued by 1,308 lenders to 4,349 unique public firms. Table 2 reports the summary statistics for various firm and loan characteristics. Detailed definitions of the variables are provided in the Appendix. All dollar values are adjusted to 2017 dollars using the Consumer Price Index data from the Bureau of Labor Statistics. To mitigate the influence of outliers, all continuous variables are winsorized at the 1st and 99th percentiles. The average loan spread in our sample is about 209 basis points, and 8% of loans are issued to borrowers whose main operations are in the neighborhood of a county hit by a natural disaster in the preceding 24 months. In our sample, an average borrowing firm has total assets of $ 8.3 billion, cash ratio of 0.09, market-to-book of 1.77, book leverage of 0.33, ROA of 0.13, PPE-to-assets ratio (Tangibility) of 0.31, and Z-score of 3.14. On average, loans have a size of around $708 million and a maturity of around 52 months. Among these loans, 40% use performance pricing and 63% are relationship lending.

TABLE 2Descriptive Statistics

TABLE 2 Long description

The table consists of 13 rows of variables and 6 columns of statistics. All variables have 32,446 observations.

* Spread (basis points): Mean 208.58, P 25 125.00, Median 200.00, P 75 275.00, Std. Dev. 119.44.

* Neighbor: Mean 0.08, P 25 0.00, Median 0.00, P 75 0.00, Std. Dev. 0.28.

* Total assets ($millions): Mean 8,258.85, P 25 484.46, Median 1,922.49, P 75 6,652.37, Std. Dev. 22,139.81.

* Cash: Mean 0.09, P 25 0.02, Median 0.05, P 75 0.13, Std. Dev. 0.11.

* Market-to-book: Mean 1.77, P 25 1.18, Median 1.49, P 75 2.02, Std. Dev. 0.94.

* Leverage: Mean 0.33, P 25 0.18, Median 0.30, P 75 0.44, Std. Dev. 0.21.

* R O A: Mean 0.13, P 25 0.09, Median 0.13, P 75 0.17, Std. Dev. 0.08.

* Tangibility: Mean 0.31, P 25 0.11, Median 0.24, P 75 0.46, Std. Dev. 0.23.

* Z-score: Mean 3.14, P 25 1.52, Median 2.60, P 75 3.97, Std. Dev. 2.75.

* Loan size ($millions): Mean 707.80, P 25 97.71, Median 312.91, P 75 795.53, Std. Dev. 1,311.38.

* Loan maturity (months): Mean 51.81, P 25 36.00, Median 60.00, P 75 60.00, Std. Dev. 22.10.

* Performance pricing: Mean 0.40, P 25 0.00, Median 0.00, P 75 1.00, Std. Dev. 0.49.

* Relationship lending: Mean 0.63, P 25 0.00, Median 1.00, P 75 1.00, Std. Dev. 0.48.

A. Baseline Results

Table 3 presents our baseline results. All t-statistics reported in parentheses are based on standard errors clustered at the county level. In column 1, we include firm and loan controls, as well as firm, bank-year, and state-year fixed effects. In column 2, we further include industry-year fixed effects. We first find that the estimated coefficients of control variables are largely consistent with prior literature. For instance, larger firms and firms with higher market-to-book ratio, lower financial leverage, and higher profitability (ROA) are associated with lower bank loan costs. More importantly, we find that the coefficient of Neighbor is positive and statistically significant at the 1% level in both columns, suggesting that within 24 months of a natural disaster, banks charge higher loan spreads to firms operating in the neighboring area of the disaster than they charge to other firms.

TABLE 3Salient Disasters and Cost of Bank Loans

TABLE 3 Long description

The table presents regression results for the dependent variable Log(Spread). The columns are numbered 1 through 4. Columns 1 and 2 use O L S, while columns 3 and 4 use a Matched Sample.

Key variable results:

* Neighbor: Coefficients are 0.064 (t = 2.893), 0.077 (t = 3.353), 0.050 (t = 2.441), and 0.072 (t = 3.518). All are statistically significant at the 1% or 5% level.

* Log(Total assets): Coefficients are approximately -0.094 across all models, all significant at the 1% level.

* Cash: Positive coefficients ranging from 0.205 to 0.311, significant in most models.

* Market-to-book: Negative coefficients ranging from -0.061 to -0.028.

* Leverage: Positive coefficients ranging from 0.421 to 0.491, all significant at the 1% level.

* R O A: Negative coefficients ranging from -0.636 to -0.847, all significant at the 1% level.

* Tangibility: Positive but statistically insignificant coefficients.

* Z-score: Negative coefficients ranging from -0.006 to -0.018.

* Log(Loan size): Negative coefficients ranging from -0.059 to -0.072, all significant at the 1% level.

* Log(Loan maturity): Statistically insignificant coefficients near zero.

* Performance pricing: Negative coefficients ranging from -0.054 to -0.070, all significant at the 1% level.

* Relationship lending: Negative coefficients ranging from -0.008 to -0.019.

Fixed Effects and Statistics:

* Industry-year F E: ‘No’ for columns 1 and 3; ‘Yes’ for columns 2 and 4.

* Number of observations: 32,561 for columns 1 and 3; 32,446 for columns 2 and 4.

* R-squared: Values are 0.844, 0.874, 0.875, and 0.902 for columns 1 through 4 respectively.

The effect on loan spreads is also economically meaningful. Estimates from column 1 indicate that if a firm mainly operates in the neighboring area of a region hit by a natural disaster in the prior 2 years, its loan spreads are 6.4% higher than those paid by other firms. Given the average loan spreads of 209 basis points (bps), this corresponds to a 13.4 basis-point higher interest rate paid by neighboring firms within 24 months of a disaster. The economic magnitude is comparable to that of market-to-book (a 1-standard-deviation reduction in market-to-book increases the cost of bank loans by 12 bps) and ROA (12.5 bps). In Supplementary Material Table A1, we include loans to firms directly affected by natural disasters and compare the impact on loan pricing for directly affected firms with the impact observed for neighboring firms. We find that both neighboring and directly affected firms experience significant increases in loan spreads following natural disasters, with magnitudes that are similar across the two groups. This comparison provides a benchmark for interpreting the economic magnitude of the pricing response observed for neighboring firms and suggests that overreaction to salient disaster events may impose meaningful financing costs on borrowing firms.

Our results are consistent with our salience hypothesis that bank lenders respond differently to neighboring firms after a disaster. To mitigate the concern that our results are driven by fundamental differences between neighboring firms after the disaster and remote firms, we repeat our baseline analysis using an entropy-balanced sample. This technique has certain advantages over the propensity-score-matching method (e.g., Chapman, Miller, and White (Reference Chapman, Miller and White2019)). It assigns weights to control observations so that the distributional properties of the treatment group and the postweighting control group are virtually identical.

Specifically, via a maximum-entropy reweighting scheme (Hainmueller and Xu (Reference Hainmueller and Xu2013)), we perform entropy balancing on the first three moments (i.e., mean, variance, and skewness) of all firm covariates to ensure that the distributions of all included control variables are similar for neighboring firms and other firms. Table 3 columns 3 and 4 present the results, which are similar to those in columns 1 and 2, suggesting that, after a natural disaster, bank lenders tend to charge higher interest rates on loans for firms operating in the neighborhood of a disaster region.

B. Dynamic Trend

To further assess causality, we examine the dynamic effects of salient natural disasters on neighboring firms’ borrowing costs. This analysis helps address concerns that the observed effect is a result of preexisting trends or spurious correlation. In addition, it allows us to evaluate whether the effect is transitory, as suggested by Dessaint and Matray (Reference Dessaint and Matray2017) and Alok, Kuma, and Wermers (Reference Alok, Kuma and Wermers2020).

Specifically, we construct a series of time indicators, each capturing whether a borrowing firm’s neighboring area has experienced (or will experience) a significant natural disaster within a given 6-month window. We then replace Neighbor with these time indicators and re-estimate the baseline specification. For example, the indicator variable Neighbor1h equals 1 if the borrowing firm’s neighboring area experienced a significant natural disaster within the past 6 months. The coefficient estimates of the time indicators for the periods before and after the disaster are plotted in Figure 1 and reported in Table 4.

FIGURE 1Dynamic Effect

Figure 1 plots the coefficient estimates along with their 90% confidence intervals in the dynamic trend analysis showing changes in loan spreads around natural disasters on a semiannual basis. The time indicators on the x-axis represent whether the borrowing firm’s neighboring area has experienced a significant natural disaster within a given 6-month period. For example, “1” indicates that the borrowing firm’s neighboring area has experienced a significant natural disaster within the past 6 months.

FIGURE 1 Long description

A line graph titled Dynamic Effect. The vertical Y-axis is labeled Coefficients and ranges from minus 0.1 to 0.2 in increments of 0.1. The horizontal X-axis is labeled Per Six-Month Relative to Disaster Date with markers at minus 3, minus 2, minus 1, 1, 2, 3, 4, and 4 plus. A vertical red line is positioned between minus 1 and 1, representing the disaster event.

Data points are represented by blue dots with dashed vertical error bars indicating 90 percent confidence intervals.

* To the left of the red line (pre-disaster):

- At minus 3, the coefficient is slightly below 0.

- At minus 2 and minus 1, the coefficients are near 0, with confidence intervals crossing the zero line.

* To the right of the red line (post-disaster):

- At 1, the coefficient rises to approximately 0.06.

- At 2, the coefficient peaks at approximately 0.12.

- At 3, 4, and 4 plus, the coefficients gradually decline toward 0.04, though they remain positive.

The confidence intervals for all post-disaster periods except 3 and 4 are entirely above the zero line, indicating a statistically significant increase in loan spreads following the disaster.

TABLE 4Dynamic Effects on Cost of Bank Loans

TABLE 4 Long description

A regression table with two columns representing different model specifications for the dependent variable Log open parenthesis Spread close parenthesis. The rows list coefficients for time-indexed variables related to natural disasters in neighboring areas, with t-statistics in parentheses below each coefficient.

* Neighbor underscore 3 h: Column 2 shows minus 0.028 with a t-statistic of minus 0.933.

* Neighbor underscore 2 h: Column 1 is 0.007 (0.215); Column 2 is 0.005 (0.168).

* Neighbor underscore 1 h: Column 1 is minus 0.001 (minus 0.046); Column 2 is 0.001 (0.032).

* Neighbor 1 h: Column 1 is 0.057* (1.691); Column 2 is 0.059* (1.730).

* Neighbor 2 h: Column 1 is 0.120*** (2.890); Column 2 is 0.121*** (2.877).

* Neighbor 3 h: Column 1 is 0.065 (1.265); Column 2 is 0.074 (1.457).

* Neighbor 3 h plus: Column 1 is 0.031 (1.321).

* Neighbor 4 h: Column 2 is 0.051 (1.279).

* Neighbor 4 h plus: Column 2 is 0.033 (1.315).

Summary statistics at the bottom show:

* Number of observations: 32,446 for both columns.

* R super 2: 0.874 for Column 1 and 0.876 for Column 2.

Asterisks indicate statistical significance: *** for 1 percent, ** for 5 percent, and * for 10 percent levels.

Two key findings emerge. First, the coefficients for predisaster periods (i.e., −3 h to −1 h) are not statistically different from 0, indicating that the increase in loan spreads does not occur prior to the disaster. Second, the postdisaster increase in loan spreads diminishes within 1 year after the disaster. These patterns support a causal interpretation of our results and align with prior evidence that the effect of salience bias decreases over time (Dessaint and Matray (Reference Dessaint and Matray2017), Alok, Kuma, and Wermers (Reference Alok, Kuma and Wermers2020)). Overall, we find no significant change in loan spreads before a neighboring disaster, but a significant increase immediately afterward.

C. The Impact of the Syndicate Structure

The loans in our sample are mostly syndicated, reflecting a collective lending arrangement where multiple banks jointly negotiate and approve terms. We next explore whether this structured collaboration affects banks’ reactions to salient disaster events. We argue that broadly syndicated loans facilitate greater group collaboration, while concentrated syndicates weaken group decision-making by centralizing power with lead arrangers—who may be more sensitive to perceived disaster risks due to reduced risk sharing. If group decision-making attenuates behavior biases, we expect a stronger salience effect in loans with more concentrated syndicate structures.

To test this conjecture, we construct three measures of syndicate concentration: the number of lenders in the syndicate, the maximum share held by the leader arranger, and the HHI of lender shares. We then divide loans to neighboring firms into two groups based on whether the number of lenders is above (i.e., Neighbor_large syndicate = 1) or below the sample median; whether the leader arranger’s maximum share is above (i.e., Neighbor_high max share = 1) or below the sample median; and whether the HHI of lender shares is above (i.e., Neighbor_high share HHI = 1) or below the sample median.

Columns 1–3 of Table 5 report the results. We find that the positive impact on loan spreads is significantly stronger when the syndicate is more concentrated, that is, when there are fewer lenders, the leader arranger retains a larger share, or the HHI of lender shares is higher. In column 4, we examine extreme cases in DealScan where loans to neighboring firms have only one lender (a nonsyndicated loan). We find consistent results. The behavior biases are far more pronounced for solo-lender loans than for syndicated loans. Taken together, Table 5 demonstrates that a concentrated lender base amplifies salience bias, whereas a broad syndicate structure helps mitigate behavior biases.

TABLE 5The Effect of the Syndicate Structure

TABLE 5 Long description

A statistical table with four columns representing different syndicate structure metrics. The dependent variable for all columns is Log(Spread).

* Column 1 (Number of Participating Lenders): Neighbor_small syndicate (A) has a coefficient of 0.111 (t-stat 4.336, significant at 1 percent). Neighbor_large syndicate (B) has a coefficient of 0.042 (t-stat 1.373). The p-value for the coefficient difference (A > B) is 0.015 (significant at 5 percent).

* Column 2 (Max Lender Shares): Neighbor_high max share (A) has a coefficient of 0.113 (t-stat 4.394, significant at 1 percent). Neighbor_low max share (B) has a coefficient of 0.045 (t-stat 1.405). The p-value for the coefficient difference (A > B) is 0.031 (significant at 5 percent).

* Column 3 (Lender Share H H I): Neighbor_high share H H I (A) has a coefficient of 0.120 (t-stat 4.360, significant at 1 percent). Neighbor_low share H H I (B) has a coefficient of 0.048 (t-stat 1.576). The p-value for the coefficient difference (A > B) is 0.023 (significant at 5 percent).

* Column 4 (Solo Lender): Neighbor_solo lender (A) has a coefficient of 0.239 (t-stat 2.131, significant at 5 percent). Neighbor_syndicate (B) has a coefficient of 0.071 (t-stat 2.888, significant at 1 percent). The p-value for the coefficient difference (A > B) is 0.081 (significant at 10 percent).

All columns report 32,446 observations with R-squared values ranging from 0.874 to 0.875.

D. Heterogeneous Effects

We next conduct several heterogeneity tests to better understand the mechanism of the impact. If it is salience bias that drives our results, we expect that the impact should be greater when the disaster events are more salient to bank officers. We use two proxies to measure the level of salience. The first one is media coverage. Greater media coverage of natural disasters likely increases the salience of the events. To test this conjecture, for each event listed in Table 1, we search for the disaster’s name in Factiva to obtain the total number of news articles related to the event within 3 months of the disaster’s occurrence. We then divide disaster-neighboring firms into two groups based on the number of news articles related to the disaster. Column 1 of Table 6 reports the results. Consistent with our conjecture, we find that the documented effect is driven by disasters with high media coverage (i.e., Neighbor_high media coverage = 1). The coefficient for the high-media-coverage subsample is positive and significant at the 1% level, and it is significantly greater than the coefficient for the low-media-coverage subsample, which is not statistically different from 0.

TABLE 6The Effect of Media Coverage and Loan Officer Location

TABLE 6 Long description

A regression table with two columns. The dependent variable for both is Log(Spread).

Column 1: Media Coverage.

- Neighbor_high media coverage (A): coefficient 0.102 with a t-statistic of 3.770, significant at the 1 percent level.

- Neighbor_low media coverage (B): coefficient 0.002 with a t-statistic of 0.069.

- p-Value for coefficient difference (A > B): 0.014, significant at the 5 percent level.

- Number of observations: 32,446.

- R-squared: 0.874.

Column 2: Loan Officer Location.

- Neighbor_close officer (A): coefficient 0.146 with a t-statistic of 2.170, significant at the 5 percent level.

- Neighbor_distant officer (B): coefficient 0.007 with a t-statistic of 0.060.

- p-Value for coefficient difference (A > B): 0.123.

- Number of observations: 9,128.

- R-squared: 0.863.

The second measure is the geographical proximity of loan officers to the disaster area. The salience of a natural disaster should be greater when loan officers are located closer to the disaster region. While certain details about loan officers (such as names, titles, employers, etc.) can be manually collected from credit agreement documents that are filed to the Security Exchange Commission (SEC), obtaining precise information about the officer’s exact location is challenging. Nevertheless, we use the residential location of loan officers in their voting registration record to estimate their proximity to the disaster (Dagostino, Gao, and Ma (Reference Dagostino, Gao and Ma2023)).Footnote ⁷ We then define two indicator variables: Neighbor_close officer, which equals 1 if the loan officer resides in the disaster state, and Neighbor_distant officer, which equals 1 if the loan officer does not reside in the disaster state. Column 2 of Table 6 reports the results. The sample size is significantly reduced due to the availability of information on loan officers’ voting registration record. We find a large and statistically significant coefficient for the variable Neighbor_close officer, but a small and statistically insignificant coefficient for Neighbor_distant officer. While the coefficient difference between these coefficients is not statistically significant at the 10% level (p = 0.123), the pattern of results is consistent with the salience hypothesis.Footnote ⁸

In Table 7, we examine the role of firms’ access to external financing. If bankers’ salience bias leads to higher borrowing costs for firms, this effect should be most pronounced among bank-dependent borrowers with limited access to alternative financing options, and consequently, weaker bargaining power. Larger firms and firms with credit ratings arguably have better access to the capital markets. Moreover, firms with a larger lender base might be able to counteract the bias of a particular bank by potentially switching to alternative banks. Therefore, we define three indicator variables to proxy for a firm’s access to alternative financing: whether the firm’s market capitalization is above the median for all neighboring firms (column 1), whether it has a credit rating (column 2), and whether its number of lenders over the past 5 years is above the median for all neighboring firms (column 3). The results indicate that increased borrowing costs among neighboring firms are indeed concentrated on smaller firms, firms lacking credit ratings, and firms with a smaller lender base. In addition, we construct an index of access to external finance for each disaster-neighboring firm by aggregating the three indicators discussed previously. We then divide neighboring firms into two groups based on the median value of this index (column 4). As shown in column 4, the impact on loan spreads is positive and statistically significant at the 1% level for firms with low access to external finance (i.e., Neighbor_low access to finance = 1), and the effect is significantly greater than for firms with high access. Overall, the findings from these heterogeneous effect analyses provide clear evidence of salience bias affecting banks’ assessment of their borrowers’ natural disaster risk.

TABLE 7The Effect of Firms’ Access to Alternative Financing

TABLE 7 Long description

A table with four columns representing different regression models for Log(Spread).

Column 1 (Firm Size):

- Neighbor_small firm (A): 0.175 (t-stat 3.700, significant at 1% level).

- Neighbor_large firm (B): 0.058 (t-stat 2.434, significant at 5% level).

- p-Value for coefficient difference (A > B): 0.006.

- Observations: 32,446. R-squared: 0.874.

Column 2 (Credit Rating):

- Neighbor_without credit rating (A): 0.107 (t-stat 3.789, significant at 1% level).

- Neighbor_with credit rating (B): 0.044 (t-stat 1.523).

- p-Value for coefficient difference (A > B): 0.029.

- Observations: 32,446. R-squared: 0.874.

Column 3 (Lender Base):

- Neighbor_small lender base (A): 0.113 (t-stat 3.786, significant at 1% level).

- Neighbor_large lender base (B): 0.060 (t-stat 1.903, significant at 10% level).

- p-Value for coefficient difference (A > B): 0.076.

- Observations: 30,884. R-squared: 0.884.

Column 4 (Index for Access to financing):

- Neighbor_low access to finance (A): 0.125 (t-stat 4.692, significant at 1% level).

- Neighbor_high access to finance (B): 0.005 (t-stat 0.133).

- p-Value for coefficient difference (A > B): 0.002.

- Observations: 30,884. R-squared: 0.885.

E. Robustness Tests

In this section, we conduct a set of tests to evaluate the robustness of our finding. The results are reported in the Supplementary Material Table A2. First, we use alternative definitions of Neighbor. In our main analysis, we define neighboring firms as those with more than 50% of sales generated from the disaster neighborhood area. Using two continuous measures of exposure—the share of sales and the share of employees in the disaster neighborhood area—we find consistent results, confirming that our results are not sensitive to the 50% sales cutoff.

Second, we remove hurricanes and the financial crisis period from our sample. As shown in Table 1, most of the natural disasters in our sample are hurricanes that predominantly affect coastal counties. We therefore repeat our baseline analysis by removing loans issued to hurricane-neighboring firms within 24 months of a hurricane. Our results remain robust. Furthermore, to mitigate the potential confounding effect of the financial crisis, we conduct an additional analysis excluding loans issued during 2007–2009. The results are similar.

Third, we conduct a package-level analysis. Multiple loan facilities can belong to the same loan package and be governed by the same contract (e.g., Bushman, Gao, Martin, and Pacelli (Reference Bushman, Gao, Martin and Pacelli2021)). We therefore repeat our baseline analysis using package–bank-level observations, where the dependent variable is the weighted average loan spread, with facility amounts used as weights. The results remain robust.

Four, we control for seasonality in the analysis. Natural disasters (e.g., hurricanes), corporate business operations, and credit markets can have seasonality (Murfin and Petersen (Reference Murfin and Petersen2016)). To account for the potential impact of seasonal factors, we include quarter fixed effects in the regressions and find similar results.

Fifth, we include firm–bank fixed effects. Unobserved stable factors, such as long-term lending relationships, could influence loan pricing. As a further robustness check, we estimate a specification with firm–bank pair fixed effects, effectively restricting identification to loan pricing changes within the same firm–bank pair over time. Our results remain robust.Footnote ⁹

Sixth, we address the issues of repeated observations by adjusting the weights. Because the number of repeated firm-facility observations increases with the number of participating banks, facilities involving more banks receive greater weight in the estimation. To address this issue, we follow prior literature (Armstrong, Jagolinzer, and Larcker (Reference Armstrong, Jagolinzer and Larcker2010), Stuart (Reference Stuart2010)) and reweight repeated observations. Specifically, we conduct a weighted least squares (WLS) analysis, assigning each observation a weight inversely proportional to its sample frequency. The results remain robust.

Finally, we ensure that our results are robust to alternative estimation methods. We adopt two additional approaches. First, we implement a DID design using our pooled loan data (e.g., Wooldridge (Reference Wooldridge2019)). Specifically, we consider the first time a borrowing firm’s neighborhood experiences a disaster within our sample period as an exogenous shock and retain loans issued to the firm 2 years before and after the disaster. For each loan issued to a treated firm, we select five loans issued to firms that: i) were never neighbors to the disaster area; ii) belonged to the same industry and year as the treated firm; and iii) had the nearest score using the propensity score matching method based on all firm covariates. We then conduct a pooled DID regression using this much smaller sample. The results reported in Supplementary Material Table A3 indicate that loan spreads are higher for borrowers in the neighborhood of a disaster area within 2 years following the disaster.

Second, we construct a firm-year panel data and use a firm’s average cost of debt as a proxy for its bank loan costs. To be included in the sample, we retain only firms that have had bank loans during our sample period. Following the literature on the cost of debt, we measure a firm’s cost of debt in year t as the ratio of its interest expenses in year t and the average total debt in years t – 1 and t. We then conduct a series of DID estimations using these panel data, with cost of debt as the dependent variable. Supplementary Material Figure A1 presents the dynamic effect of natural disasters on neighboring firms’ cost of debt using the two-way fixed effects (TWFE) estimator and four alternative dynamic DID estimators developed by de Chaisemartin and D’Haultfoeuille (Reference De Chaisemartin and D’Haultfoeuille2020), Callaway and Sant’Anna (Reference Callaway and Sant’Anna2021), Sun and Abraham (Reference Sun and Abraham2021), and Borusyak, Jaravel, and Spiess (Reference Borusyak, Jaravel and Spiess2024)). Similarly, there is no significant difference in the cost of debt prior to the occurrence of a natural disaster. The treatment effect mainly concentrates during and within the first year following the disaster year. However, we are cautious about drawing strong inferences due to the limitations of this coarse measure of bank loan costs in these analyses.

Overall, the findings from these analyses lend further support to our main argument that disasters in neighboring areas affect firms’ cost of debt in the subsequent year and mitigate the concern that this effect is driven by omitted variables.

A. Alternative Explanations

In this section, we evaluate potential alternative explanations that may confound our interpretation. Specifically, we discuss the following possible interpretations: regional spillovers, limited credit supply, and lender rent extraction.

B. Impact on Loan Allocation Decisions

Our analysis thus far shows that, conditioning on lending activity, banks charge higher spreads on neighboring firms in the period following natural disasters, and that these effects are unlikely to be driven by firm, bank, or local economic fundamentals. Building on these findings, we next explore banks’ responses at the extensive margin. Specifically, we examine whether banks temporarily adjust their lending portfolios by reducing exposure to firms located in neighboring areas following salient natural disasters.

To answer this question, we conduct a bank-county-level analysis in which the dependent variable is the amount of new lending by a bank to firms in a given county (as recorded in DealScan), scaled by the bank’s total new loan issuance in that year. The results are reported in Table 8. All regressions include state-year, bank-year, and bank-county fixed effects, allowing us to compare changes in the fraction of lending by the same bank to the same county over time while controlling for time-varying factors at the state and bank levels.

TABLE 8Impact on Loan Allocation Decisions

TABLE 8 Long description

A statistical table with three columns labeled 1, 2, and 3, representing different regression models for the dependent variable percent Bank County Allocation.

Key variables and results include:

* Neighbor underscore 2 h: Model 2 is 0.004 (0.414); Model 3 is 0.004 (0.377).

* Neighbor underscore 1 h: Model 2 is 0.002 (0.239); Model 3 is minus 0.001 (minus 0.110).

* Neighbor county: Model 1 is minus 0.010* (minus 1.873); Model 2 is minus 0.009* (minus 1.809).

* Neighbor 1 h: Model 3 is minus 0.014* (minus 1.917).

* Neighbor 2 h: Model 3 is minus 0.018** (minus 2.364).

* Neighbor 3 h: Model 3 is minus 0.005 (minus 0.483).

* Neighbor 4 h: Model 3 is minus 0.009 (minus 1.160).

* Neighbor 4 h plus: Model 3 is 0.006 (0.745).

* Income: Model 1 is 0.005 (1.374); Model 2 is 0.005 (1.373); Model 3 is 0.004 (1.385).

* Unemployment rate: Model 1 is minus 0.001 (minus 0.131); Model 2 is minus 0.001 (minus 0.127); Model 3 is minus 0.000 (minus 0.032).

Summary statistics at the bottom show the Number of observations is 15,050 and the R-squared is 0.800 for all three models. Asterisks indicate statistical significance levels.

In column 1, we find a negative and statistically significant coefficient on Neighbor county, indicating that banks allocate a smaller amount of new lending to firms in neighboring counties following a natural disaster. In columns 2 and 3, we include time indicators that capture whether the county has experienced (or will experience) a significant natural disaster within a given 6-month window. The results show that the negative impact on lending does not occur prior to the disaster and is concentrated in the 12 months afterward. Overall, these results suggest a temporary reallocation of bank lending following natural disasters, providing corroborating evidence of a similar reaction at the extensive margin that is consistent with salience-driven overreaction in the aftermath of such events. These findings are analogous to the portfolio reallocation effects documented by Alok, Kuma, and Wermers (Reference Alok, Kuma and Wermers2020) for mutual fund managers’ equity holdings in regions affected by major disasters.

C. Additional Analysis