Does IRS Monitoring Matter for the Cost of Bank Loans?

Table 5 presents the baseline estimations regarding H.1. Our models show a good fit with a 54% adjusted R² on average. In model 1 of Table 5, we control only for the effect of Audit rate_t-1 on the cost of bank loans; while in model 2 of Table 5, we also account for loan characteristics; and in model 3, we add firm-level controls as well. The coefficients for Audit rate _t-1 are significant at the 1% level and negative (-0.021, -0.022, -0.020) in models 1, 2, and 3 of Table 5, respectively.

Overall, these results support H.1 that firms' bank loan spreads decrease when IRS monitoring is more stringent. Based on the average of the three coefficients of the lagged IRS audit rate in the first three models of Table 5 (-0.021), our findings, economically, indicate that one standard deviation increase in the lagged IRS audit rate (8.651) leads to an 8.84 basis points decrease in bank loan spreads (8.834 = \({e}^{0.021}*8.651)\). Another way to assess the relevance of these findings is reflected by the interest savings based on the average loan size in our sample of $522 million, and the average time to maturity of around 3.4 years. As per our estimates, a one standard deviation increase in Audit rate_t-1 means around $1.49 million in interest savings (1.49 = 522*0.00084*3.4). Additionally, we show that larger firms with more tangible assets, higher profitability, lower risk, lower tax avoidance, less leverage, higher rating, and listed on the S&P enjoy lower borrowing costs, which is in line with similar studies (Graham et al. 2008; Bharath et al. 2011; Hasan et al. 2014; Huang et al. 2020). Furthermore, large loans with a shorter maturity are associated with lower loan spreads. These findings are consistent with Chava and Roberts (2008) who show that banks exposed to the risk of a longer maturity charge higher loan spreads as compensation.

The findings from the baseline models in Table 5 show that the lagged IRS audit rates and the cost of bank loans are negatively related. These results support H.1. This support is due to the ability of stringent IRS monitoring to decrease information asymmetry.

The baseline models 1–3 in Table 5 assume that the IRS audit rate variable is exogenous to the cost of bank loans measure. We try to ease endogeneity concerns in our baseline models by using lagged IRS audit rates. To further address the potential endogeneity between the IRS audit rate and the cost of bank loans, we proceed with a two-stage least squares instrumental variable (2SLS-IV) estimation. In the first stage, we obtain predicted values of the IRS Audit rate t-1 using an OLS regression that comprises two instruments and all the control variables we employ in our baseline analysis. In the second stage, we replace the IRS audit rate variable with its predicted values that we obtain from the first stage.

For the first instrument, we follow the extant literature (Hoopes et al. 2012) and use the natural logarithm of the number of corporate tax returns in each asset size class filed in the previous year (Ln tax returns of size class t-2). By previous year we mean with respect to the Audit rate t-1. Hence, this instrument has a two-year lag from the dependent variable (i.e., the cost of bank loans). The IRS audit rate represents the number of actual corporate returns audited by the IRS in each firm size class in a given year divided by the number of corporate tax returns filed in each firm asset size class in the previous year. Consequently, IRS audit rates depend on fluctuations of corporate tax returns filed across time in the same firm size group. The numerator component (i.e., the number of corporate tax returns audited by the IRS) represents the IRS audit effort in each asset size class. In the case of a feedback effect, firms could, at least in theory, attempt to influence the IRS audit effort in their asset size class. Hence, the numerator component of the IRS audit measure, which represents the IRS audit effort, is more susceptible to endogeneity concerns. The denominator component (i.e., the number of corporate tax returns filed in each asset size class in the previous year), which we use as an instrument following Hoopes et al. (2012), is less susceptible to reverse causality because it is not likely that a firm will try to influence its IRS audit probability rate by not filing a tax return. Furthermore, since this instrument is not a measure of the IRS audit effort, it is unlikely that it would affect the cost of bank loans in a way other than its influence on the IRS audit rate for each asset size class (i.e., the exclusion restriction). We expect a negative and significant correlation of this instrument with the IRS audit rate in the first stage of the IV estimation.

We rely on the intuition that the IRS faces budget constraints each year for the second instrument. Such constraints mean that the IRS has limited audit resources to allocate to each asset size class each year. The IRS, in different periods, shifts its resources to specific target groups for auditing purposes (Scholz and Wood 1998; Bagchi 2016), such as, in our case, specific firm size classes. This action constraints the resources that the IRS could dedicate to monitoring and auditing firms in other asset-size groups. The shift of resources of the IRS to specific target groups for auditing purposes is also compatible with anecdotal evidence from the IRS operations. For example, the Transactional Records Access Clearinghouse (TRAC) states in 2008, '…By ordering its revenue agents to concentrate on the smaller corporations that normally take a lot less time to audit, the agency (i.e., the IRS) was able to push up the overall number of corporate audits…'. Based on the above discussion, we construct a second instrument by assigning to each borrowing firm in a given asset size class the lagged average IRS audit rate of the rest of the asset size classes (Average IRS audit rate of other size classes t-1). We expect this instrument to have a negative and significant relationship with the IRS audit rate of each asset size class. Furthermore, this instrument should, in principle, not directly affect the cost of bank loans for firms in a given asset size class since it does not represent the IRS audit rate for this specific group of firms (i.e., the exclusion restriction).

In model 4 of Table 5, we use the 2SLS-IV. The first-stage results show that the first instrument (Ln tax returns of size class _t-2) has a significant association with the IRS audit rate at the 1% level and has the expected (i.e., negative) coefficient sign (see the lower part of model 4 in Table 5). Also, the first-stage findings show that the second instrument (Average IRS audit rate of other size classes _t-1) displays a negative and significant relationship with the IRS audit rate at the 1% level (see the lower part of model 5 in Table 5). In model 6 of Table 5, we run 2SLS estimations using both instruments. We find that our two instruments continue to have a negative and significant association with the IRS audit rate at the 1% level. These results show that both instruments fulfil the inclusion criterion.

Furthermore, in model 6 of Table 5, the validity of the instruments is supported by the under-identification LM test (UIT), the weak identification Wald F-Test (WIT) using critical values from Stock and Yogo (2005), and the Hansen over-identification test of Hansen (OIT). The second stage results (models 4–6 of Table 5) show that the predicted values of the IRS audit rate (Pred Audit rate _t-1) exert a negative and significant effect on the cost of bank loans at the 1% level. These findings further support H.1.

We use three alternative proxies of lending relationships to test H.2A and H.2B: RIA, RIN, and REL DUM. Table 6 depicts the results.

In models 1, 3, and 5, we first run the baseline specification with all three variables. They have a negative but non-significant coefficient in these models. This finding is consistent with other studies (e.g., Hagendorff et al. 2021). A potential explanation is that the potential savings in screening and monitoring could be offset by the extraction of informational rents (Hasan et al. 2019). This is rational since both hold-up effects and savings in screening and monitoring costs coexist in lending relationships (Alexandre et al. 2014).

Models 2, 4, and 6 in Table 6 comprise the interactions between the three variables (RIA, RIN, and REL DUM) and Audit rate _t-1. In models 2 and 4, we find that the interactions between Audit rate _t-1 and RIA and RIN have negative and significant coefficients at the 1% level. In the same models, the individual effect of Audit rate t-1 on loan spreads is also negative and significant at the 1% level. We find similar results when using REL DUM in model 6 of Table 6.

The positive and significant individual effect of the lending relationship variables (RIA, RIN, and REL DUM) denotes that when IRS monitoring is weak, banks are more able to extract informational rents from their relationships with borrowers. However, as IRS monitoring becomes stringent, relationship banks' informational advantage vis-à-vis other providers of external financing diminishes. Consequently, banks are less able to hold-up their relationship borrowers as the negative interaction between the lending relationships variables and the IRS monitoring proxy indicates.

To visually demonstrate the findings that support hypothesis H.2B, we offer two interaction plots of RIA and RIN with Audit rate _t-1. These interaction plots in Figs. 1 and 2 illustrate that the negative association between IRS monitoring and loan spreads becomes more apparent when lending relationships intensify.

The empirical evidence supporting H.2B is consistent with studies that identify the factors that decrease information asymmetries and improve firms’ information environment and therefore increase borrowers' bargaining power and decrease banks' ability to extract informational rents from their lending relationships (Schenone 2010; Cahn et al. 2021; Saidi and Žaldokas 2021). In our context, the ability of IRS monitoring to decrease information asymmetry could weaken the informational advantage of incumbent banks and render it more credible that relationship borrowers could access alternative providers of external financing (Guedhami and Pittman 2008; El Ghoul et al. 2011; Bauer et al. 2021; Gallemore and Jacob 2020). Hence, banks may perceive that their relationship borrowers have increased bargaining power in the presence of more stringent IRS monitoring.

Next, we investigate the type of borrowers for which the findings for H.2B are more evident. The public markets for external financing rely more strongly than the private lending market on external monitoring from entities such as the IRS. Investors in the public market have weaker monitoring incentives than banks because they are usually more widely dispersed (Diamond 1984; Boot and Thakor 2009; Bharath and Hertzel 2019). Furthermore, the public markets depend more strongly than banks on high-quality publicly available hard information (Liberti and Petersen 2019). Banks could supplement this information with inside (private) information on the borrowing firm that public investors do not have access to (Fama 1985; Dass and Massa 2011). Thus, the public markets are likely to more highly value the ability of the IRS to decrease information asymmetry than private lenders. Therefore, IRS monitoring might be more important for decreasing the informational advantage, and hence the capacity to extract rents from the lending relationships of incumbent banks vis-à-vis the public markets than other private lenders.

We use three variables to represent a firm's ability to access the public markets for external financing. The first variable indicates if a firm is listed in the stock market (Listed). The second is the variable that captures if a firm is a member of the prestigious S&P 500 index (S&P500). The third variable denotes if a firm has an investment-grade credit rating (INV GRADE). Then, we use models that comprise triple interaction terms between these three proxies and Audit rate _t-1 and RIA, RIN, and REL DUM.

Table 7 depicts the findings. In Panel A of Table 7, we run models that estimate the effect of the triple interaction term on the cost of bank loans. Panel A shows that the triple interaction is negative and significant across all models (1–3), at least at the 10% level. In Panel B of Table 7, we repeat this exercise using the S&P500 dichotomous variable as one of the regressors in the triple interaction. We observe similar findings in models 1–3 where we use the three alternative measures of lending relationships. The triple interactions are negative and significant, at least at the 5% level. Lastly, we run similar estimations using INV GRADE. Model 3 in Panel C shows that the triple interactions are negative and significant at the 1% level like our earlier findings.

These findings support our theoretical prediction that under more stringent IRS monitoring, the cost of bank loans decreases more for relationship borrowers that have credible access to the public markets of external financing. By decreasing information asymmetry, IRS monitoring reduces incumbent banks’ informational advantage and facilitates less costly access to the public markets (Allen and Gottesman 2006; Santos and Winton 2008; Hale and Santos 2009; Saunders and Steffen 2011). Therefore, the borrowers' bargaining power vis-à-vis banks could increase. As a consequence, the ability of incumbent banks to extract informational rents from lending relationships diminishes.

Next, we perform a split sample analysis based on the median of RIA and RIN. This exercise provides further insights into the conditioning effect of lending relationships on the association between IRS monitoring and the cost of bank loans. The results are in Table 8. We show that the interactions between RIA and RIN and Audit rate _t-1 are negative and significant for both subsamples (see models 1–2 of Panel A and 3–4 of panel B of Table 8). These findings support H.2B.

The different results for the two subsamples reflect that the individual effect of Audit rate _t-1 on loan spreads is negative and significant only for low lending relationships (compare models 1 and 2 with models 3 and 4 in Panel A of Table 8). We interpret this result in the following way: When lending relationships are weaker, the ability of the IRS monitoring to decrease information asymmetry could be of some value to the banks' lending. Weaker lending relationships mean that banks have not yet collected a high level of private information on borrowers. Therefore, IRS monitoring could be helpful for the screening and monitoring of borrowers at this stage. For example, the quality of borrowers' information environment, such as the financial reporting transparency that IRS monitoring enhances, could be more useful when banks hold less proprietary information and are less familiar with borrowers (Bharath et al. 2008; Berger et al. 2017).

However, as lending relationships strengthen, the negative association between Audit rate _t-1 and loan spreads becomes more evident through its negative and significant interaction with RIA and RIN. This finding indicates that at a higher level of lending relationships, IRS monitoring increases borrowers' bargaining power and restrains banks' ability to extract rents from the private information they accumulate on borrowers. Indeed, for the high lending relationship subsample (see models 1 and 2 of Panel B of Table 8), the negative association of IRS monitoring with loan spreads occurs only via the negative and significant interaction with the lending relationship variables (i.e., H.2B). The insignificant coefficient for Audit rate _t-1 in the high lending relationship subsample further indicates that in this context, the low information asymmetry stemming from past interactions renders IRS monitoring less beneficial for the relationship banks' screening and monitoring operations.

Another difference between the lower and higher lending relationship subsamples is that the coefficients for RIA and RIN are different; the one for the first is positive and significant but the one for the second is not significant. A potential explanation is the confluence of the distinct hold-up and the pass-through savings effects of screening and monitoring that coexist in lending relationships (Alexandre et al. 2014). After gaining a new borrower, banks tend to increase loan prices that is a behavior consistent with the hold-up effect (Ioannidou and Ongena 2010). When lending relationships intensify to a high level, the information asymmetry between borrowers and banks reaches a minimum. Consequently, substantial savings in screening and monitoring could offset the hold-up effect.

We continue the split sample analysis with models that comprise the triple interaction terms between Audit rate _t-1; RIA and RIN; and Listed, S&P500, and INV GRADE. The results from these tests are available in Table 9.

For the high lending relationships subsample, we find negative and significant triple interaction terms between Audit rate _t-1, RIA and RIN, and the Listed and S&P500 dummies (see models 3 and 4 in Panels A and B of Table 9). The interaction term between the lending relationship measures and IRS monitoring in the same models loses significance. These findings show that credible access to external public financing facilitates the negative association between IRS monitoring and the cost of bank loans for firms that display strong lending relationships with their banks (i.e., H.2B).

The results for the low lending relationship subsample show that the negative interactions between Audit rate _t-1 and RIA and RIN remain negative and significant. However, the triple interaction terms with Listed, S&P500, and INV GRADE are not significantly different from zero. These findings show that for the low lending relationship subsample, the negative interaction between IRS monitoring and lending relationships (i.e., H.2B) is apparent for both bank-dependent firms and firms with credible access to external public financing.