The Structuring of Financial Covenants When Lenders Acquire Soft Information Robert Prilmeier∗ The Ohio State University May 14, 2011 Abstract Financial covenants aid monitoring by allocating state-contingent control rights between borrowers and lenders based on hard information. Much of the finance literature suggests that lenders’ acquisition of soft information over the course of a lending relationship supports effective monitoring. This paper investigates how the presence of soft information affects the number and tightness of hard information-driven covenants included in the contract. Consistent with borrowers trading off monitoring benefits with covenant-created hold-up costs, I find that the effect of relationship intensity on the number of covenants included in the contract follows an inverted U shape. This effect is stronger for borrowers with easier access to public debt markets and for sole lender and single lead arranger loans. I further argue that a covenant need not be particularly tight to act as a monitoring incentive. Consistent with covenant tightness addressing information asymmetry concerns, I find that tightness is reduced over the course of a relationship, especially when borrowers are informationally opaque. ∗ PhD candidate at The Ohio State University; e-mail: prilmeier 1@fisher.osu.edu 1 1 Introduction What determines the structuring of financial covenants in debt contracts? The theoretical and empirical finance literature suggests that covenants are used to reduce agency costs between debtholders and shareholders (Myers, 1977; Smith and Warner, 1979), to aid in monitoring the borrower (Chava and Roberts, 2008; Roberts and Sufi, 2009; Nini et al., 2009, 2010), and to give the lender an incentive to monitor in the first place (Rajan and Winton, 1995; Park, 2000). Consistent with banks’ special role as monitors (Diamond, 1984, 1991; Fama, 1985), bank loans contain more extensive sets of financial covenants than public debt, and covenant violations are almost always related to bank debt (Kahan and Tuckman, 1993; Rauh and Sufi, 2010). However, little is known about the structuring of financial covenants when financial institutions have varying degrees of knowledge about a borrower’s “soft” information (information not verifiable by outsiders). Financial covenants are restrictions in debt contracts that are written on “hard” information, i.e. accounting quantities that are verifiable in court. If the restriction stipulated by a covenant is violated, the lender obtains the right to demand immediate repayment of the loan. However, firms may violate covenants even when their future prospects are good. Indeed, many covenant violations are waived (Beneish and Press, 1995; Chen and Wei, 1993; Dichev and Skinner, 2002; Roberts and Sufi, 2009). Thus, it is likely that lenders use soft information to determine the optimal response to a covenant violation. In addition, soft information creates information asymmetries between borrowers and lenders and between lenders that have a relationship with the borrower and those that do not. How, then, does lenders’ acquisition of soft information in a banking relationship affect the structuring of financial covenants when contracting a loan? Finance theory offers an array of predictions regarding this question. While the agency theory of covenants of Myers (1977) and Smith and Warner (1979) suggests that covenants create value by reducing agency costs, they increase the incidence of renegotiations between the borrower and the lender. To the extent that the acquisition of soft information in a banking relationship reduces renegotiation costs (Boot, 2000), bank relationships support the inclusion of a broader set of covenants which will lead to renegotiation more frequently. Rajan and Winton (1995) and Park (2000) develop models in which the lender’s claim competes with the claims of other stakeholders, such as investors, trade creditors or employees who are able to free-ride on 2 the lenders’ monitoring effort. This reduces the lender’s incentives to monitor and shut down bad projects that the manager cannot credibly commit to shutting down. Rajan and Winton (1995) show that in this case, covenants can be used as an incentive to monitor since they make the effective maturity of the loan contingent on monitoring. In addition, Park (2000) shows that it is optimal to assign the monitoring incentive to the lender with the lowest monitoring cost. Since the acquisition of soft information in a relationship is likely to reduce the cost of monitoring (Boot, 2000), loans from relationship lenders should include a more extensive set of covenants. However, the presence of information asymmetries between the borrower and the lender predicts the opposite. Such information asymmetries are largest when the two parties have no prior relationship. To the extent that covenants are used to protect the lender from borrower moral hazard, one would expect the need for covenants to decrease as the lender learns the true nature of the borrower (Boot, 2000). When a lending relationship is sufficiently exclusive, the lender may be able to develop an information monopoly about the borrower in order to extract rents when renewing the loan (Greenbaum et al., 1989; Sharpe, 1990; Rajan, 1992). Such rent extraction could include imposing an overly extensive set of financial covenants on the borrower. However, financial covenants are unique among loan contract terms in the sense that they themselves may create a hold-up problem. When a borrower violates a financial covenant, the relationship lender is likely to be better informed than potential outside lenders about the true prospects of the firm. If the efficient decision is to liquidate the borrower, this makes no difference, but if the borrower’s prospects are good, the violation will allow the relationship lender to extract rents. Provided that borrowers have sufficient bargaining power when entering the loan agreement, they should seek to reduce the intensity of financial covenants attached to the loan. These theories offer conflicting predictions. It is important to note that they are not mutually exclusive and they may interact with each other. For example, a borrower who is being held up by his lender at the contracting point will likely be unable to negotiate a reduced covenant load in order to avoid a covenant-created hold-up problem. Moreover, the relationship effect may be nonlinear, although the direction of such nonlinearity is an empirical question. If borrowers determine the covenant load that maximizes firm value by trading off monitoring 3 incentives with covenant-created hold-up costs, the relationship effect should follow an inverted U-shape. However, if contracting choices are determined by a trade-off between the reduction of information asymmetries between the borrower and the relationship lender and an increase in information advantages of the relationship lender over non-relationship lenders that fosters hold-up at the point of contracting the loan, the effect is more likely to follow a U shape. I test these theories using a sample of syndicated loans to non-financial U.S. borrowers. I measure relationship intensity as the proportion of the firm’s borrowings over the previous five years that it has borrowed from the current lead arranger. I define covenant intensity as the number of financial covenants attached to a loan. Consistent with the idea that reduced renegotiation costs favor the inclusion of more restrictive covenants, but are ultimately countervailed by covenant-created hold-up problems, I find that financial covenant intensity increases with relationship intensity for low levels of relationship intensity, but decreases as the relationship becomes more and more exclusive. These results are robust to alternative measures of financial covenant intensity and relationship intensity. Further, I investigate the effect of variation in the borrower’s bargaining power and the lender’s stake in the loan. Access to the public debt market and other capital markets should increase the borrower’s ex ante bargaining power and thus enable them to negotiate away the covenant-created hold-up potential more effectively. The evidence supports this hypothesis. The decrease in financial covenant intensity in exclusive relationships is concentrated in borrowers with access to outside sources of capital. In addition, a loan’s covenant intensity affects its effective maturity relative to other creditors’ claims, but does not prevent loan participants or multiple lead arrangers from free-riding on each other. Hence, if covenants are used as a monitoring incentive, they should perform this role more effectively for sole lender loans or loans with only one lead arranger. Indeed, I find that covenant use increases more strongly in relationship intensity for such loans. Ex post competition from participant lenders or other lead arrangers in the loan syndicate at the time of a covenant violation might alleviate the covenantcreated hold-up problem, provided that these lenders also gather some soft information. I find some evidence that the decrease in covenant intensity is concentrated in loans with one lead arranger. I do not find evidence that covenant intensity is driven by a reduction in information asymmetries between the borrower and the relationship lender. 4 The choice to borrow from a relationship lender is likely endogenous. To rule out that the results are driven by selection on observable or unobservable firm characteristics, I investigate the relationship effect on loan contract terms to which the monitoring incentive and renegotiation cost as well as the covenant-created hold-up theories do not apply. I also employ propensity score matching methods and instrumental variables estimation. The results from these methods suggest that the relationship effect on covenant intensity is causal. Loan contracts not only involve a choice of how many covenants are included in the loan, but also how tight they are. I define covenant tightness as the probability that the loan’s most restrictive covenant will be violated. Predictions for covenant tightness are not necessarily the same as for covenant intensity. Covenant tightness may be less important than covenant intensity in incentivizing lenders to monitor. In the model of Rajan and Winton (1995), the lender monitors because the covenant allows her to shorten the effective maturity of the loan when the borrower does poorly and monitoring is a precondition to exercising this right. To implement this, the covenant needs to give control to the lender in bad states of nature, but it need not necessarily be particularly tight. In addition, the borrower’s risk of being held up by the lender when violating a covenant increases in covenant intensity, but is unlikely to increase in tightness. Covenant tightness is hard information known to outside lenders. If a very tight covenant is violated, this is unlikely to be taken as a negative signal by outside lenders and hence creates little possibility for the inside lender to hold up the borrower. Demiroglu and James (2010) find that the violation of tight covenants has little impact on the borrower. Consequently, I find evidence that covenant tightness is driven by information asymmetries. After accounting for the endogeneity induced by the private information content of covenant tightness (Demiroglu and James, 2010), tightness decreases in relationship intensity, and more so in situations where a reduction in information asymmetry is likely to be important. My paper contributes to the literature on several dimensions. To the best of my knowledge, the finance literature provides little evidence on the effect of lending relationships and soft information acquisition on hard-information monitoring tools such as financial covenants. In particular, this study is the first to provide evidence of a hold-up problem created by financial covenants.1 Secondly, in a recent paper, Schenone (2010) finds a hold-up effect of banking 1 A few recent studies touch on specific aspects of banking relationships and financial covenants. Murfin (2010) finds that a borrower’s default on a loan causes the lender to tighten the covenants of loans to other borrowers, and more so if the other borrowers’ relationship with the bank is more exclusive. Ivashina and Kovner (2010) 5 relationships on yield spreads only among borrowers not listed on a stock exchange, but finds no such effect among publicly listed borrowers. The strong rights that a covenant violation confers to the firm’s creditors enable me to show that hold-up considerations apply even to publicly listed and rated borrowers and that borrowers adapt their loan contracts to trade off this cost with monitoring benefits. Finally, I provide further evidence that covenant intensity and covenant tightness are used in different ways, consistent with Demiroglu and James (2010) who find that covenant tightness contains private information about the firm’s prospects, but covenant intensity does not. The remainder of this paper is structured as follows. Section 2 describes the predictions from finance theory. Section 3 details the data collection process. Section 4 discusses the results for covenant intensity, and section 5 addresses endogeneity concerns. Section 6 presents the results for covenant tightness, section 7 performs additional robustness checks, and section 8 concludes. 2 Theory One of the primary functions of banks is the monitoring of borrowers (Diamond, 1984, 1991; Fama, 1985). Debt covenants enhance firm value by allowing control rights to shift from shareholders to debtholders when the firm is performing poorly, even outside of bankruptcy (Aghion and Bolton, 1992; Dewatripont and Tirole, 1994; Dichev and Skinner, 2002). Smith (1993) and Sridhar and Magee (1996) argue that financial covenants serve as tripwires that enable flexible monitoring, where creditors’ response can range from waivers to new restrictions. However, loan renegotiation following a covenant violation involves costs. Creditors need to assess the reasons for the covenant violation and negotiate a response with the borrower. The case may need to be negotiated in court if the borrower and the lender cannot come to an agreement. Boot (2000) argues that the soft information acquired in a banking relationship reduces such renegotiation costs and thus supports the use of covenants. While covenants can serve as monitoring tools, Rajan and Winton (1995) and Park (2000) develop models in which covenants provide the lender with an incentive to monitor in the first place. In these models, the lender’s claim competes with the claims of other investors and creditors, who will be able to free-ride on the lender’s monitoring function. This reduces the lender’s show that private equity firms with stronger bank relationships enjoy a less restrictive maximum debt to EBITDA covenant when financing a leveraged buyout. 6 incentive to acquire information about the borrower. At the same time, the entrepreneur’s inability to credibly commit to abandoning projects in bad states of nature reduces firm value. Covenants help to overcome both problems since they make the lender’s payoff contingent on continuous monitoring. A covenant breach allows the lender to reassess the borrower’s credit risk and to impose restrictions that increase firm value,2 but only if she can prove that the covenant has indeed been violated. While imposing firm value enhancing restrictions benefits other claim holders as well, being able to renegotiate in bad states of nature increases the bank’s payoff contingent on monitoring. This enhances the lender’s incentives to monitor. In the model of Rajan and Winton (1995), the lender will monitor if the value she receives from gathering the information exceeds monitoring costs. Since soft information acquired in a lending relationship lowers the lender’s monitoring costs (Boot, 2000), covenants are more likely to achieve monitoring in the case of a relationship lender. In Park’s model, the optimal debt contract involves a two-tiered structure: Monitoring is delegated to a senior lender whose claim is large enough to be impaired in the case of liquidation, so that junior lenders receive nothing in case of a liquidation and thus have no incentive to monitor. Therefore, the senior lender is given all the covenants. It then becomes optimal to assign the monitoring task to the lender with the lowest monitoring cost. To the extent that relationships lower monitoring costs, the relationship lender is likely to be in that position.3 While the above theories predict an increase in covenant use as a lending relationship progresses, the agency theory of covenants developed by Jensen and Meckling (1976), Myers (1977), and Smith and Warner (1979) predicts the opposite. According to this theory, shareholders can take a number of actions that hurt debtholders’ claims, such as risk-shifting, excessive dividend payouts, over- and underinvestment and so forth. This moral hazard necessitates the need to monitor, especially when information asymmetries are severe. Therefore, when a bank lends to a borrower it has never dealt with before, it might use a relatively extensive set of covenants to 2 For some of the measures taken by lenders after covenant violations, see Chava and Roberts (2008), Nini et al. (2009), and Nini et al. (2010). 3 A different way of implementing monitoring, of course, is the short-term debt contract. Bharath et al. (2009) find that for borrowers with low credit quality, debt maturity decreases in relationship intensity. They argue that lower monitoring costs allow lenders to shorten the maturity as a means of commitment to monitoring. However, a short maturity is not a perfect substitute for restrictive covenants. First, covenants typically are monitored at a higher rate than the frequency with which short-term debt is rolled over. Compliance reports for debt covenants are often filed on a quarterly basis (Chava and Roberts, 2008), whereas short-term bank debt is typically issued on a lower frequency. Second, Rajan and Winton (1995) show that there are situations where short-term debt without covenants cannot implement monitoring, but long-term debt with covenants can. 7 curb moral hazard. With repeated interaction, however, information asymmetries are reduced as the bank learns borrower-specific information (Boot, 2000), and it can reduce the number of covenants written into the contract. Compared to other monitoring tools, the unique feature of financial covenants is that they shorten the effective maturity of the loan conditional on a signal of poor performance. This has interesting implications for the hold-up problem in a lending relationship. With covenants, there are two different types of hold-up problems: hold-up can occur at the point of contracting the loan, or at the point at which the covenant is violated. Hold-up problems that occur when contracting a new loan or rolling over a matured loan have received a significant amount of attention both in the theoretical and the empirical literature, although to my knowledge they have not been studied specifically for covenants. The idea follows naturally from the above argument. As relationship lenders acquire borrower-specific information, the borrower-lender information asymmetry declines, but the informational advantage of the relationship lender over outside lenders intensifies. The relationship lender’s information monopoly may thus cause the borrower to become “informationally captured” (Greenbaum et al., 1989; Sharpe, 1990; Rajan, 1992). This may enable the lender to extract rents by imposing unfavorable contract terms (such as a more extensive set of restrictive covenants) on the borrower when entering into subsequent loan agreements.4 However, there is an interesting and previously unexplored twist to the hold-up problem when loan contracts contain covenants. When the firm violates a covenant, the lender gains the right to accelerate the loan and thus shorten the effective maturity. In the Rajan (1992) model, the hold-up problem arises from short-term bank debt. Since covenants shorten the maturity in some cases, but not others, one might think that their hold-up potential is somewhere in between short- and long-term bank debt. However, the nature of debt covenants suggests that their hold-up potential may be stronger. With short-term bank debt, the borrower’s loan is rolled over without prejudice when it matures. If a well-performing borrower is informationally captured by its bank, uninformed lenders will pool that borrower with less well-performing 4 Empirical studies of contracting hold-up thus far have focused on the yield spread the borrower is required to pay as the main aspect of rent extraction. A number of studies find that small, unlisted borrowers pay higher interest rates as the banking relationship progresses (Degryse and Cayseele, 2000; Ioannidou and Ongena, 2010; Schenone, 2010), although some studies find no such effect (Petersen and Rajan, 1994; Berger and Udell, 1995). Borrowers who are listed on a stock exchange have been found to pay lower yield spreads to relationship lenders (Bharath et al., 2009; Schenone, 2010). 8 borrowers and thus the relationship lender can extract rents. When a covenant is violated, however, this is a signal that the borrower may be in trouble, although he need not be given that many covenant violations are waived. Due to ongoing monitoring, relationship lenders are better able to assess the information content of the covenant violation than uninformed lenders. Outside lenders will now pool the firm with the set of violators, who are worse performers on average than the non-violators as long as covenants are written on mildly informative accounting quantities. Consider a firm that happens to violate a covenant due to a random negative realization in its accounting ratio, but whose soft information suggests that its prospects are good. The violation will leave the relationship lender’s willingness to lend unchanged, whereas it may cause outside lenders to perceive the firm to be riskier than it is. This creates a potential for rent extraction by the relationship lender, which would not exist had the covenant not been written into the contract.5 If the borrower foresees this problem when he enters into the loan agreement, one would expect him to try to negotiate it away ex ante, especially if the relationship with the lender is exclusive and hence her soft information advantage over outside lenders is large. The above theories are by no means mutually exclusive and they may be interacted with each other. For example, the borrower’s ability to avoid covenant-created hold-up problems by reducing the set of covenants included in the loan is likely to depend on his ex ante bargaining power. To the extent that a borrower is already being held up when entering the loan agreement, he will be unable to bargain for a reduction in financial covenant intensity. Thus, one should expect that the reduction in covenant intensity for exclusive relationships is contingent on the borrower’s access to other sources of capital, such as the public bond market.6 Moreover, hold-up problems are likely to be nonlinearly increasing in the exclusivity of the borrower’s relationship with the lender. To see this, suppose that X% of a borrower’s loans over the past five years were supplied by the lender with whom the current loan agreement is made.7 This means that the remaining percentage of loans were provided by other lenders who 5 Consistent with this, Roberts and Sufi (2009) find that few borrowers switch lenders after a covenant violation, even though the violation leads to an increase in interest rates and a tightening of credit on average. However, whether these effects of the violation simply reflect efficient responses on the part of the lenders or whether there is a hold-up component to them remains an open question. 6 Note that access to the public bond market would be expected to improve the borrower’s ex ante bargaining power but not necessarily his bargaining power at the point of a covenant violation since public bond market participants are likely to be even more uninformed about the borrower than outside banks. 7 Following Bharath et al. (2009), I assume that soft information acquisition stops if borrower and lender have 9 therefore have some degree of knowledge about the borrower’s soft information. As Schenone (2010) argues, lenders are likely to become increasingly effective in processing firm-specific information as the relationship progresses. If relationship intensity with the current lender increases, say, from 0% to 20%, this is unlikely to create hold-up potential since it means that the current lender has just started to learn firm-specific information and since there are other lenders who have more experience with the borrower. On the other hand, if it increases from 80% to 100%, the percentage change is the same but the impact is likely to be larger since the firm moves from a situation where some other lender is present and is processing firm-specific information to a situation where no other inside lender exists. Consequently, it is conceivable that hold-up issues dominate when relationship intensity is relatively high, whereas for lower levels of relationship intensity the other theories may be more important. This leads to the possibility of a non-linear relationship effect. In particular, if borrowers maximize firm value through an optimal level of monitoring by trading off monitoring incentives with covenantcreated hold-up costs, we should expect the effect of relationship intensity on covenant use to follow an inverted U, at least for borrowers that have sufficient bargaining power to negotiate for fewer covenants even when the relationship is exclusive. Alternatively, covenant choice may be driven by information asymmetries between the borrower and his lender and information advantages of the current lender over outside lenders, respectively. In this case, the reduction in borrower-lender information asymmetries due to a relationship should reduce the incidence of covenants initially, but the increasing information advantage of the relationship lender over outside lenders should increase the relationship lender’s ability to hold up the borrower when contractig a loan by including more restrictive covenants. This theory predicts a U shape. Schenone (2010) finds evidence in favor of this pattern for yield spreads, but only before a firm’s IPO. This suggests that publicly listed firms are able to leave or credibly threaten to leave the relationship before burdensome terms are imposed on them when contracting a new loan or rolling over an old one. However, an important difference between yield spreads and covenants is that with covenants, hold-up is state-contingent. not had a lending contact within five years. This appears reasonable since 84% of all loans in the sample have a maturity of five years or less. 10 3 The Data I obtain data on syndicated and large sole lender loans from Loan Pricing Corporation’s DealScan database. DealScan reports yield spreads, covenants, maturities and other characteristics for loans made by bank and non-bank lenders to both U.S. and foreign corporations and accounts for a large proportion of the U.S. private loan market.8 According to DealScan officials, the vast majority of information on covenants is collected from loan documents filed with the SEC. Consequently, DealScan contains little information about covenants for loans contracted before 1995, when companies started filing SEC documents electronically. Therefore, the sample ranges from January 1995 to December 2008. The sample consists of U.S. currency denominated loans obtained by U.S. firms that are not a member of the financial, utility, or public administration sectors. I merge this dataset with the borrowers’ accounting data in Compustat for the fiscal year prior to loan inception using a link file kindly provided by Michael Roberts and Sudheer Chava.9 Borrowers’ S&P long-term issuer ratings are taken at the month before loan inception to reflect the borrower’s risk assessment at the time the loan is made. After applying all filters, the final sample for which the required information is available consists of 7,923 loans incurred by 3,169 borrowers. Loans are reported in DealScan as packages (or deals), which contain one or more facilities. Information such as yield spreads, loan amounts, and maturities are available at the facility level, whereas covenants are reported at the package level. To avoid artificially weighting covenant observations by the number of facilities in the package, I aggregate all data to the package level.10 Numerous bank mergers and acquisitions occurred during the sample period. To account for the M&A activity, I matched the DealScan lenders to FDIC institution IDs (RSSD IDs) based on name, geographical location and time. I performed this match at the individual bank level rather than the bank holding company level on the theory that information acquisition primarily occurs through direct interaction. Using this match, the Federal Reserve’s National 8 According to Carey and Hrycay (1999), DealScan covers between 50 and 75% of all commercial loans (by value) in the U.S. in the early 1990s and coverage further increases thereafter. 9 Details on the construction of this link file can be found in Chava and Roberts (2008). 10 One might wonder to what extent the results are influenced by firms incurring multiple loans within one year. It turns out that in the final sample, 91.6% of the firm-year combinations are unique, while for 7.7% of the firm years there are two loans in the sample for 0.7% of the firm years there are three or four loans in the sample. Consequently, aggregating observations to firm-years does not change results. 11 Information Center allows me to track bank mergers over time and attribute an acquired bank’s relationships to the surviving entity. I measure both financial covenant and non-financial covenant intensity as count variables that add one for each financial and non-financial covenant, respectively, as recorded in DealScan. Table 1 details the various types of financial and non-financial covenants. Financial covenants are grouped into six categories: debt to balance sheet, coverage, debt to cash flow, liquidity, net worth, and EBITDA covenants. Debt to balance sheet and debt to cash flow covenants restrict the maximum indebtedness the borrower is allowed to incur relative to the various balance sheet and cash flow items detailed in table 1. Coverage, liquidity, net worth, and EBITDA covenants all require the maintenance of certain minimum coverage or liquidity ratios or of a minimum net worth or EBITDA. Among financial covenants, coverage covenants are the most frequent, with 79% of all loans containing at least one coverage covenant. Debt to cash flow covenants and net worth covenants are included in 60% and 43% of the loans, respectively. Non-financial covenants include sweep provisions, dividend restrictions and capital expenditure restrictions. Sweep provisions require the borrowing firm to repay part or all of the loan prematurely if it takes certain actions. For example, if a loan carries an asset sales sweep, the borrowing firm must use asset sale proceeds in excess of certain allowances to repay the loan. Close to 80% of all loans have a dividend restriction, while about one fifth of the loans have a capital expenditure restriction. Among the 38% of all loans that carry a sweep provision, the majority has more than one such provision.11 The empirical predictions require measuring relationship intensity in a way that captures both the lender’s prior experience with the borrower as well as the exclusiveness of the relationship. I define the lender as the loan’s lead arranger since the lead arranger acts as an intermediary between the borrower and the participant lenders and thus is better informed (Ivashina, 2009; Guerin, 2007). I designate as lead arrangers any lender for which the field “lead arranger credit” is denoted as “Yes” in DealScan as well as the lenders of all sole lender loans. In addition, I search the field “lender roles” and define the following roles as lead arrangers: 11 Note that data on non-financial covenants in DealScan is sometimes missing, even when data on financial covenants is available. Since the vast majority of DealScan’s covenant information comes from loan documents filed with the SEC, whenever there is information on financial covenants, information on all covenants included in the loan should be available to LPC. Thus, I set non-financial covenants to zero if the information is missing, but data on financial covenants is available. This method appears to be similar to practitioners’ approach (e.g. May and Verde (2006)). In any case, results for financial covenants are unaffected by this issue. 12 agent, administrative agent, arranger, lead bank. These definitions coincide with Bharath et al. (2009), who describe these roles in more detail. I then identify all instances in DealScan where the borrower obtained funding in the five years prior to the current loan (including loans that are not in the final sample due to missing information) and measure relationship intensity as follows:12 P j Relation (Max Amt) = max Loan amountj ∗ I(k) , j Loan amountj P k (1) where I(k) indicates lead arranger k’s participation in loan j. In words, I determine relationship intensity as the total amount of loans over the past five years for which the current lead arranger acted as a lead arranger divided by the total amount of all loans over the past five years. If there are no loans in the previous five years, the measure is undefined. Thus, I require at least one prior loan to be observable. If the current loan has more than one lead arranger, I take the maximum of that ratio across all lead arrangers since soft information-driven decisions are likely to be led by the lead arranger that has learned the most soft information. As an alternative, I consider a relationship intensity measure that gives equal credit to each lead arranger involved in the loan and calculates the sum of relationship intensities across lead arrangers: P Relation (Sum Amt) = j X k Loan amountj /Nj ∗ I(k) P , j Loan amountj (2) where Nj indicates the number of lead arrangers participating in loan j. Relative to the Relation (Max Amt) measure, this measure has the advantage that it equals one if and only if there is no lender outside of the current syndicate who has ever acted as a lead arranger to the firm. The disadvantage is that it is not clear why a lender in a syndicate with two lead arrangers should learn only half the information that a lender in a syndicate with one lead arranger would learn. In any case, the two measures differ only if multiple lead arrangers are involved in any of the firm’s loans.13 Consequently, they have a correlation of 0.957 and lead to very similar results. Table 2 shows the number of loans available per firm for the final sample and for the sample used to determine relationship intensity. The median firm has two loans in the final sample and four loans in the sample used to determine relationship status. Thus, using all available loans 12 This measure is also used by Bharath et al. (2009) and Schenone (2010). Among the loans in the sample, 74.4% have one lead arranger, 22.1% have two lead arrangers and 3.5% have more than two lead arrangers. 13 13 in DealScan to calculate relationship intensity is important if one wants to be able to detect any nonlinear effect in relationship intensity. Section 7 will discuss the reasons why these loans are excluded from the final sample and perform robustness checks to ensure that this does not affect the results. Table 3 presents univariate tests of differences in firm and loan characteristics conditional on relationship intensity. Relationship intensity is categorized as “low” if Relation (Amt) is less than 30%, “high” if it is more than 70%, and “medium” if it is in between. The results show that financial covenant use increases by about 4% from the low to the medium category and decreases by about 7% from the medium to the high category. The changes are highly statistically significant. Non-financial covenant use essentially does not change from the low to the medium category and decreases by about 13% from the medium to the high category, again statistically significant. Yield spread and collateral requirements decrease in relationship intensity, whereas maturity is largely unchanged. However, table 3 also shows that firm characteristics are correlated with relationship intensity, with firms with high relationship intensities being larger, more likely to be a member of the S&P 500, more likely to be rated, having better ratings, and having a lower current ratio. Therefore, I now turn to multiple regressions.14 4 Multiple Regressions for Covenant Intensity 4.1 Baseline Tests Since the number of financial covenants is a count variable, I test the effect of relationship intensity on financial covenant intensity by estimating Poisson regressions. As argued in section 2, this effect may be nonlinear. One way to test for nonlinearity is to assume a quadratic form: log(FinCovi ) = α1 + β1 Relationi + γ1 Relation2i + δ1 Controlsi + 1,i , (3) where FinCov is the number of financial covenants included in a deal and Relation is one of the two relationship intensity measures. Since the linear and squared terms of relationship intensity 14 In unreported regressions, I confirm the negative effect of relationship intensity on spreads, collateral requirements, and maturity found by Bharath et al. (2009) and obtain very similar coefficients. Although data availability on financial covenants is not as extensive as for the terms they investigate, this similarity in results lends support to the quality of my sample. 14 are highly correlated, one may be concerned that regression estimates are an artifact of this correlation. Therefore, I focus on presenting results using a dummy variable specification: log(FinCovi ) = α2 + β2 LowRelationi + γ2 HighRelationi + δ2 Controlsi + 2,i , (4) where Low Relation and High Relation are dummy variables that equal one if relationship intensity is below 30% and at least 70%, respectively. Consequently, loans with medium relationship intensities become the base group, which allows testing for the existence of any U-shape. Controls include various loan and firm characteristics as displayed in table 3 as well as industry fixed effects at the one-digit SIC industry level, year fixed effects, and loan purpose and loan type fixed effects. If one deal contains two different types of loans, e.g. a revolver and a term loan, then both these dummy variables equal one for that deal.15 As is customary in studies using Compustat data, the top and bottom 1% of all financial ratios are winsorized to control for outliers.16 Table 4 shows the results. The effect of relationship intensity on financial covenant intensity appears to follow an inverted U-shape. For the quadratic specifications, the linear term is significantly positive and the quadratic term significantly negative. For the dummy variable specifications, both the low relationship and high relationship dummies indicate a significantly lower covenant intensity compared to loans with a medium relationship intensity. Table 4 also shows that financial covenant use decreases in the size of the loan and in firm size. The coefficient of leverage is positive as expected, but not significant. Covenant use increases in the number of lenders participating in the loan, which mirrors the result in Drucker and Puri (2009) that loans sold on the secondary market contain more covenants than loans that banks keep on their balance sheet. Both the current ratio and the coverage ratio enter positively in the regression. Many covenants are written on ratios related to these two. Such covenants may be more informative if these ratios are above a certain threshold.17 Firms with a worse credit 15 I exclude loan purpose and loan type dummies that have fewer than ten nonzero observations in the sample. These winsorizations have an effect on the coefficients of some of these ratios, but they do not change the results regarding the relationship effect. 17 Current ratio covenants typically stipulate a minimum ratio of 1.0 or higher, while interest coverage ratio covenants typically stipulate a minimum of 1.25 or higher. If one excludes loans from borrowers whose ratios are below one of these thresholds, the current ratio and coverage ratio are no longer significant in the regressions. 16 15 rating or no credit rating at all are subject to more covenants, while loans to members of the S&P 500 carry fewer covenants.18 The relationship effect is economically significant. Figure 1 plots the effect of relationship intensity on covenant intensity using the quadratic specification from regression (2) in table 4 and a stepwise dummy variable specification using the same controls (not reported in the table). Financial covenant intensity increases by about 8% from low to medium relationship intensity and decreases by about 4% from medium to high relationship intensity. These changes are equivalent to the effect of a change in the rating by two to three notches and by one to two notches, respectively. For further comparison, a one standard deviation increase in leverage leads to an increase in financial covenant intensity by about 1% (or 2% if one removes the rating variable from the regression) and a one standard deviation increase in the log of assets leads to a decrease by about 4% (or 7% if one removes the S&P 500 dummy from the regression). The relationship effect on covenants is also similar in size to the effect on other loan terms. For example, Bharath et al. (2009) find that a change in relationship intensity from 0% to 100% leads to a decrease in the all-in-spread drawn by 5% (evaluated at their sample average spread). The stepwise specification in figure 1 also shows that the relationship intensity thresholds of 30% and 70% used for the dummy variable specification in regression (3), while somewhat arbitrary, capture the curve quite well. A variety of alternative cutoffs exists that would yield similar or stronger results. The results in this section show that the effect of relationship intensity on financial covenant use follows an inverted U. I now turn to determining which of the theories described in section 2 contribute to this effect. 4.2 Bargaining Power and Syndicate Structure An important way of distinguishing hold-up problems from information asymmetry effects is to consider the borrower’s bargaining power. If financial covenant violations provide an exclusive lender with an opportunity to hold up the borrower, borrowers would be expected to seek to avoid this hold-up potential by writing fewer covenants into their loans. However, to be able 18 A potential problem with Poisson models is over- or underdispersion of the data relative to the model. Calculating the deviance for the models in table 4 and dividing by the degrees of freedom gives a value of 0.324, which is smaller than one and hence suggests that underdispersion is present. Standard errors scaled by the square root of the deviance-based dispersion are slightly smaller than standard errors clustered at the firm level. 16 to do so, they are likely to need sufficient bargaining power at the time the loan agreement is entered into. Thus, the hold-up argument predicts that a borrower with better access to outside capital will be better able to adjust its loan contracts for the covenant violation holdup problem by writing fewer covenants when the relationship with the current lender is more exclusive. Because borrowers with access to outside capital markets tend to be more transparent, an information asymmetry story would predict the opposite: The decrease in covenant intensity for high relationship intensities should be more pronounced for opaque firms that do not have access to outside capital markets. I use the existence of an S&P long-term issuer rating as well as a firm’s access to the commercial paper market (as evidenced by a short-term rating of A-2 or better (Murfin, 2010)) as proxies for the firm’s access to public debt markets.19 Columns 1 and 5 of table 5 show that the decrease in covenant intensity for firms in exclusive relationships is concentrated in rated firms. Among rated firms, covenant intensity is between 6% and 7% lower for borrowers in exclusive relationships as compared to borrowers in medium intensity relationships.20 Figure 2 plots the difference in the relationship effect for rated vs. unrated firms using a quadratic specification (not reported in table 5). For rated firms, covenant intensity increases until a relationship intensity of about 53% and decreases strongly thereafter, while for unrated firms covenant intensity increases until a relationship intensity of about 60% and remains relatively constant for higher relationship intensities.21 Columns 2 and 6 of table 5 show that the interaction between high relationship intensity and access to the commercial paper market has similar coefficients to the interaction with being rated, but is at best marginally significant statistically. Consequently, it appears that having any rating is more important than having a rating that indicates a particularly high credit quality. I next turn to the impact of syndicate structure on the relationship effect. This matters 19 One might wonder whether the hold-up problem itself is smaller for rated borrowers. However, note that investors on the public debt market are most likely even more uninformed than potential outside lenders on the market for bank debt. Thus, access to the public debt market provides the borrower with bargaining power ex ante, but not necessarily ex post after a covenant violation has occurred. In any case, this would hurt my identification strategy. 20 This effect does not appear to be driven by the fact that controlling for the level of the rating helps the regression model better measure credit quality for rated borrowers than for unrated borrowers. When dropping the ordinal rating variable and thus leveling the playing field, coefficients on the interaction terms remain qualitatively and quantitatively similar. 21 Ai and Norton (2003) point out that the interpretation of interaction terms can be difficult in nonlinear models. This problem does not apply here since the Poisson model is linear in the log of the covenant count, which makes analyzing percentage changes straightforward. Couching the discussion in terms of percentage changes appears reasonable since financial covenants are written on correlated accounting ratios. On average, adding a second covenant should increase restrictiveness more than adding a tenth covenant. 17 for two reasons. First, an increase in the number of participants limits the extent to which monitoring benefits accrue to the lead arranger. In the Rajan and Winton (1995) model, covenants incentivize the lender to monitor despite the presence of other claim holders because a covenant violation allows the lender to demand early repayment or adjustments to the interest rate, the benefits of which accrue solely to her since other claim holders do not hold these rights. In a borrowing syndicate with multiple loan participants or multiple lead arrangers, every lender is treated equally in the event of a covenant breach, which again allows other lenders to free-ride on the monitor. Consequently, if monitoring incentives motivate the inclusion of financial covenants in contracts with relationship lenders, the increase in covenant use from low to medium relationship intensities should be stronger for sole lender loans than for loans syndicated to other loan participants and it should be stronger for loans with one lead arranger than for loans with multiple lead arrangers. Second, to the extent that other loan participants, or especially, co-lead arrangers learn soft information about the borrower, the potential to hold up a borrower when he violates a covenant may be reduced. If a hold-up attempt occurs, an informed competitor within the syndicate could offer the borrower better terms and win his business. This incentive to deviate for other syndicate members would limit the syndicate’s ability to hold up the borrower in the first place. If this is the case, the need to reduce covenant intensity for high relationship intensities would be reduced. Columns 3 and 7 of table 5 show the increase in covenant intensity with an increase from low to medium relationship intensity is stronger for sole lender loans, consistent with covenants providing monitoring incentives or sole lenders investing more strongly in the acquisition of soft information. According to columns 4 and 8, the inverted U curve is flatter for loans with multiple lead arrangers, although the difference is significant only for the relationship measure that gives 1/N credit to each of the N lead arrangers. This is consistent with both monitoring incentives and within-syndicate competition among lead arrangers.22 22 One may wonder whether loans with multiple lead arrangers are purely transactional in nature, such that no soft information is acquired. In section 5.1, I show that yield spreads decrease in relationship intensity by the same amount for loans with single vs. multiple lead arrangers. Hence, soft information acquisition does seem to occur in both types of loans. 18 4.3 The Borrower’s Information Opacity As described in section 2, a reduction in information asymmetries over the course of a banking relationship might result in a lower need for financial covenants. I now test whether this is the case and whether the reduction in covenant intensity for exclusive relationships is driven by such an effect. Proxies for information opacity include dummy variables indicating whether the borrower’s total assets were below the sample median during the start year of the loan, whether the borrower’s stock was a member of the S&P 500 index, whether the borrower is a high tech firm (following Loughran and Ritter (2004)), whether the number of analysts following the borrower’s stock was below the sample median for that year, whether the dispersion of analyst forecasts for the borrower’s earnings per share is above the median, and whether the borrower is listed on NASDAQ as opposed to NYSE or Amex. It should be noted that many of these proxies are negatively related to a firm’s access to capital markets and, hence, its ex ante bargaining power. To the extent that bargaining power matters more than information asymmetries, one would expect results to mirror those of the previous section. Although opaque firms might be easier to hold up after a covenant violation, they are likely to lack the bargaining power to adapt the loan contract to this hold up potential in the first place. Table 6 presents results using the Relation (Max Amt) measure. Results using the Relation (Sum Amt) measure are qualitatively and quantitatively similar and are omitted for brevity. Table 6 shows that the evidence is inconsistent with a lower need for covenants due to a reduction of information asymmetries over the course of a relationship. The downward sloping part of the inverted U is stronger for large borrowers, borrowers with a large analyst following and borrowers whose stock is part of the S&P 500. The first two of these interactions are statistically significant at the 5% level, while the interaction with S&P 500 membership is marginally significant at the 10% level. There is no difference for high tech vs. other firms, NASDAQ vs. NYSE/Amex firms, or firms with high vs. low dispersion of analyst forecasts, proxies that arguably are less related to capital markets access and more related to pure information asymmetries. In addition, there is no significant difference in upward slopes for low relationship intensities across borrowers of different opacity. Taken together, the results presented thus far support the theory that soft information acquisition in a banking relationship enhances covenant use due to a reduction in renegotiation 19 and monitoring costs and that borrowers trade off this benefit with the hold-up potential arising from a covenant violation. The evidence does not support the theory that the initial upward slope in covenant intensity is driven by ex ante hold-up since the initial increase is also present (or stronger, if anything) for large, and rated firms which succeed in negotiating a lower financial covenant intensity even in exclusive relationships. Likewise, the evidence is inconsistent with covenant intensity being driven by a reduction in information asymmetry between the borrower and the relationship lender since I find that the reduction in covenant intensity in exclusive relationships is driven by large, rated firms rather than small, opaque firms. 5 Endogeneity The choice of forming, developing, and breaking a banking relationship is likely to be endogenous. Firms that do not form relationships might differ from firms that have relationships with several banks and firms that have an exclusive relationship in ways that explain the inverted U-effect documented thus far. Note that any such endogeneity story would also have to account for the finding that the reduction in covenant intensity is concentrated in relatively transparent firms that have higher ex ante bargaining power. While it seems difficult to construct such a story, this section employs three different ways to test whether results are driven by selection on observable or unobservable firm characteristics. The first strategy analyzes loan terms to which neither the monitoring incentive argument nor the covenant-created hold-up argument applies. There should not be an inverted U-effect for these other loan terms. The second strategy discusses relationship effects estimated by propensity score matching methods and the third uses an instrumental variables approach. 5.1 Relationship effects on yield spreads and non-financial covenants Yield spreads do not offer the state-contingent control feature embedded in financial covenants. For this reason, neither the monitoring incentive and renegotiation cost argument nor the argument that a covenant violation creates hold-up is applicable to yield spreads. The two other theories presented in section 2 do apply to yield spreads. If relationship lenders succeed in holding up their borrowers at the contracting point, yield spreads should increase in relationship intensity. If relationships mitigate information asymmetries between the borrower and the 20 relationship lender, yield spreads should decrease in relationship intensity. Given these theories, finding an inverted U-effect of relationship intensity on yield spreads would call into question the conclusions drawn above. It would mean that the inverted U could be driven by sample selection effects or, since yield spreads are related to credit risk, by failing to control for an important risk factor. Bharath et al. (2009) study the effect of relationship intensity on yield spreads for a sample of borrowers contained in DealScan and find that yields spreads decrease in relationship intensity, and more so for informationally opaque borrowers, consistent with the information asymmetry theories. Nevertheless, it appears worthwhile to perform a similar analysis for my sample since their sample selection criteria differ from mine23 and since they do not allow for nonlinearity of the relationship effect. In table 7, I regress the all-in spread drawn provided by the DealScan database on relationship intensity and the same controls that are used in the previous regressions. The relationship intensity measure used is the Relation (Max Amt) measure defined in equation 1, but conclusions are not affected by using the Relation (Sum Amt) measure. Column (1) shows that, consistent with Bharath et al. (2009), yield spreads decrease in relationship intensity in my sample. Column (2) allows for a potential inverted U shape, but does not find such an effect. Yield spreads on loans with low relationship intensity are insignificantly larger than those on loans with medium relationship intensity and yield spreads on loans with high relationship intensity are significantly smaller. I further analyze the effect of the relationship asymmetry and bargaining power proxies. As in Bharath et al. (2009), columns (3) and (4) show that the decrease of yield spreads in relationship intensity is driven by unrated and small borrowers who are likely more opaque. The result is exactly the opposite of what I find for financial covenants. This reduces the likelihood that the documented decrease of financial covenant intensity in exclusive relationships for large, rated borrowers is due to unobserved risk. Columns (5) and (6) show that the relationship effect on yield spreads is not significantly larger for borrowers listed on NASDAQ, but it is significantly larger for borrowers with higher dispersion of analysts’ earnings forecasts. Importantly, column (7) does not reject the hypothesis that the effect of relationship intensity on yield spreads for loans with multiple lead arrangers is the same as the effect for loans with one lead arranger. This supports the interpretation that the weaker 23 Most notably, I require the availability of information on financial covenant intensity and I analyze loans at the deal level, while they focus on the facility level. 21 inverted U-shape for loans with multiple lead arrangers is due to loan-level financial covenants’ inability to overcome monitoring incentive problems and due to within-syndicate competition at the covenant violation point. It is inconsistent with the argument that loans with multiple lead arrangers are simply transactional loans and no acquisition of soft information occurs. A further test in table 7 concerns non-financial covenants. While financial covenants define certain financial health criteria that the borrower might fail to meet for a plethora of reasons, non-financial covenants restrict specific actions that typically involve some form of moral hazard, such as capital expenditures, dividend payouts, and asset sales. Financial covenants are thus written on quantities that are more volatile and less directly controllable by managers than those that non-financial covenants contract on (Kahan and Tuckman, 1993). This means that financial covenants require a larger effort on the part of the monitor to evaluate the implications of a violation and that they are less transparent to uninformed lenders. Consequently, monitoring incentive, renegotiation cost and violation hold-up considerations should be more pronounced for financial covenants, whereas non-financial covenants should be more strongly related to information asymmetries. Columns (8) and (9) in table 7 test this prediction. Allowing for a linear term shows a significantly negative effect of relationship intensity on non-financial covenants. When allowing for non-linearity, there is no evidence of an inverted U-shape for non-financial covenants. The difference between medium and low relationship intensity loans is insignificant,24 whereas high relationship intensity loans carry significantly fewer non-financial covenants. This is again inconsistent with the interpretation that the inverted U-effect of relationship intensity on financial covenant use is driven by a background factor that affects the choice of loan terms in general. 5.2 Propensity score matching While the results using yield spreads and non-financial covenants are instructive and corroborate the interpretation of the results presented in section 4, I now turn to direct strategies of accounting for endogeneity. Ideally, one would like to randomly assign firms to groups that are treated with either low, medium, or high relationship intensity and then observe their covenant choices. In reality, however, firms are not assigned to treatment groups at random, and one 24 While the sign is negative using the Relation (Max Amt) measure, it is positive and still insignificant using the Relation (Sum Amt) measure. 22 cannot observe what outcome e.g. a firm choosing a relationship loan would have experienced had it chosen a non-relationship loan. However, for each firm that receives treatment, one can try to find untreated firms that ex ante had the same likelihood to receive treatment and estimate the average treatment effect on the treated (ATT) as the average of the difference in financial covenant intensity between the matches. This can be done using propensity score matching (PSM) as described in Rosenbaum and Rubin (1983) and Heckman et al. (1997, 1998). Assuming that all factors affecting selection into treatment groups are observable, the resulting estimate of the ATT is unbiased. Selection on unobservables, however, cannot be cured with PSM and requires the use of an instrumental variables approach (see section 5.3). I implement the PSM technique using the following steps. First, I estimate each firm’s probability to be assigned to the low, medium, or high relationship intensity group using an ordered probit model where the dependent variable equals one for low, two for medium, and three for high relationship intensity, respectively. The ordered probit model uses firm and loan characteristics from table 4 with year, industry, loan purpose and loan type fixed effects as independent variables. I exclude the maturity of the loan and syndicate size from the regression to address concerns that these may be endogenous themselves. Results including these variables are qualitatively and quantitatively similar. The ordered probit model yields three propensity scores – one for each relationship category. Next, I estimate the ATT for medium relationship borrowers by matching each borrower with medium relationship intensity to a set of borrowers with low relationship intensity based on each borrower’s propensity to be treated with medium relationship intensity. In the same fashion, the ATT for high relationship intensity borrowers is obtained by matching them with medium relationship intensity borrowers based on the propensity to be in the high intensity group. There are many ways to implement the matching estimator. For increased comparability and transparency, I use the same estimators that Bharath et al. (2009) use.25 Nearest neighbor estimators calculate the difference in financial covenant intensity between each treated loan and the arithmetic average for the n loans in the untreated group with the closest propensity scores. Following Bharath et al. (2009), I report results using n = 10 and n = 50. Kernel estimators construct a counterfactual using a weighted average of financial covenant intensities for loans of untreated firms. Weights decrease in the propensity score difference between the treated and the 25 I implement all estimators using the Stata module PSMATCH2 provided by Leuven and Sianesi (2003). 23 untreated firms. The Gaussian kernel uses all untreated observations as matches, but weights decrease faster the lower the bandwidth chosen to estimate the kernel. The Epanechnikov kernel only uses untreated observations within the propensity score bandwidth. I employ a bandwidth of 0.01. For the nearest neighbor matching, I estimate standard errors using the AbadieImbens (2006) variance estimator. Standard errors for the kernel estimators are obtained from bootstrapping with 1,000 replications.26 Table 8 shows the results for the PSM technique. Column (1) matches all medium relationship intensity loans with low relationship intensity loans and finds an ATT for financial covenant intensity of about 0.13, or 5.1% relative to the sample average financial covenant intensity of 2.54. This effect is slightly smaller than in the Poisson model, but highly statistically significant. Column (4) matches all high relationship intensity loans to medium relationship intensity loans, yielding a statistically significant ATT of about -0.09, or -3.5% relative to the sample average, in line with the results from the Poisson model. As Bharath et al. (2009) point out, there are two distinct groups of non-relationship borrowers: those that never form relationships and those that just broke up an existing relationship to borrow from a non-relationship lender. These groups may differ from each other in financial covenant use e.g. because borrowers break up relationships when a new lender offers them particularly favorable contract terms or, to the contrary, because the new lender needs to guard against an adverse selection problem (Detragiache et al., 2000). Following Bharath et al. (2009), I split the sample of low relationship intensity loans into two groups. The first group, which becomes the untreated group in column (2) of table 8, consists of loans where the previous loan had a relationship intensity larger than zero, but the current loan is a non-relationship loan, thus breaking up an existing relationship. The second group, used as the untreated group in column (3) of table 8, consists of loans that do not represent such a break up of an existing relationship.27 It turns out that the increase in financial covenant intensity when moving from 26 Estimating standard errors for PSM poses some challenges. Abadie and Imbens (2008) show that the bootstrap is not valid for nearest neighbor matching estimators due to a lack of smoothness. The variance estimator in Abadie and Imbens (2006) is asymptotically consistent assuming that propensity scores are known. In practice, however, propensity scores are estimated. Interestingly, Abadie and Imbens (2009) show for the variance of the average treatment effect (ATE) that adjusting for the estimation of propensity scores in the first step reduces the asymptotic variance of the estimator. While it is not clear whether this also applies to the variance of the ATT reported here, note that results for the ATE are virtually identical. To my knowledge, it is not clear whether the kernel estimators are smooth enough for the bootstrap to be valid. However, conclusions are unaffected by instead using the unconditional variance estimator provided in Lechner (2001). 27 If the previous loan has an undefined relationship status (because it is that borrower’s first loan in DealScan), this determination cannot be made and the loan is excluded from this part of the analysis. 24 low to medium relationship intensity is statistically significant regardless of comparison group, although coefficients are somewhat larger when using firms that just broke up a relationship. In a similar fashion, high relationship intensity borrowers can be distinguished into two groups: those that always borrow from the same lead arranger and those that have at least once in the past five years borrowed from some other lead arranger. It may be the case that the first group is completely locked into their relationship, whereas the second group has the ability to borrow from other lead arrangers, but chooses not to do so. Columns (5) and (6) address this concern by restricting the treatment group to loans with a relationship intensity of exactly one (column (5)), and a relationship intensity of more than 0.7, but less than one (column (6)). As table 8 shows, this distinction does not matter. Coefficients are highly similar across the two columns and the high relationship intensity effect is statistically significant in both. 5.3 Instrumental variables Selection into a particular relationship status may be driven by unobservable factors such as the firm’s and lender’s private information about the future prospects of the firm. Perhaps firms with good unobservable credit quality tend to either borrow only transactional loans (never forming a relationship) or focus on a relationship with one bank, while firms with poor unobservable credit quality maintain relationships with more than one bank, thus creating the observed inverted U-effect of relationship intensity. From a theoretical standpoint, it is not clear why this should be the case. Nor is it consistent with the findings for yield spreads and the fact that the inverted U is concentrated in loans obtained by rated firms, for which the information on credit quality that is available to the researcher is more precise, if anything. Nevertheless, such a concern can be addressed using instrumental variables (IV) estimation. The key to IV estimation is to find an instrument that is correlated with relationship status, but has no effect on financial covenant intensity other than through relationship status. Since this study uses two endogenous variables — low and high relationship intensity indicators —, at least two instruments are needed to identify both endogenous variables. Bharath et al. (2009) use the distance between the borrower and the lead arranger as an instrument. They argue that geographical proximity fosters the gathering and processing of soft information and hence the formation of relationships, while not affecting loan terms per se.28 I use this instrument as well. 28 See their paper for a discussion of the literature on geography and banking relationships. 25 Historical addresses (city, state, and ZIP code) of borrowers’ headquarters are obtained from the header of the corresponding 10-K filing using DirectEDGAR.29 Historical lender addresses are from Call Reports and the National Information Center (NIC) of the Federal Reserve System, which means that foreign lenders and non-bank lenders not covered by the NIC are excluded from this analysis. I translate these data into geographical longitudes and latitudes using the WebGIS application provided by the University of Southern California.30 I then compute the log of one plus the spherical distance in miles between the borrower and the lead arranger using the formula given in Dass and Massa (2010).31 A similar geographical argument can be made when the borrower and the lender are located in the same state. In this case, the lender’s familiarity with the state-level economic, legal, and political environment is likely to positively affect the processing of soft information in a relationship. At the same time, this proximity should not affect loan terms other than through its effect on relationship formation. Hence, as a second instrument, I use an indicator variable that equals one if at least one lead arranger is located in the same state as the borrower. One concern with using two geographical instruments is that they may be too close in meaning to be able to identify two distinct endogenous variables. For this reason, I also employ two proxies for the unconditional expectation of relationship formation in the borrower’s industry. Older and more established industries in which the average firm is relatively large are likely to be more transparent and have access to a wider variety of capital sources. Consequently, banking relationships are less likely to be exclusive. I proxy for this using the median size of a firm in the borrower’s industry in the year prior to the loan start date as well as industry age, which I define as 2008 minus the year of the earliest Compustat IPO date of any firm in the borrower’s industry.32 These variables are likely to be correlated with relationship status of firms in the borrower’s industry, but should not affect the loan terms of that particular borrower.33 Since the endogenous regressors represent a nonlinear function of the same variable, the potential of instrument weakness is an immediate concern. To my knowledge, no procedure is available to test for instrument weakness in the nonlinear generalized method of moments 29 I thank Burch Kealey for help with this data. At the time of writing, this service is available at https://webgis.usc.edu/Default.aspx. 31 If the loan has more than one lead arranger, I take the distance between the borrower and the closest lead arranger on the theory that the group of lead arrangers as a whole will be at least as informed as the most well-informed lead arranger. 32 I group firms into industries using the Fama-French 38 industry classification. 33 The industry median size instrument is also used in Dass and Massa (2010). 30 26 setting, whereas tests and implications in the linear model are better understood (see Stock et al. (2002)). For this reason, I implement a two-stage least squares (2SLS) estimation of equation 4.34 Because Low Relation and High Relation are indicator variables, I follow the recommendation in Wooldridge (2002) to first estimate a probability model for the relationship dummies with the instruments described above and then use the predicted probabilities as the actual instruments in the first stage of the 2SLS estimation.35 To account for the dependence between the indicator variables, I choose an ordered probit model, but results using two separate binary probit models are similar. Results from the IV estimation are displayed in table 9.36 I present three models to assess whether results depend on the choice of instruments. All four instruments are statistically significant predictors in the first stage ordered probit. The coefficients on the instrumented relationship dummies remain negative and statistically significant, but their magnitudes are very large compared to the Poisson regressions. Two instrument weakness tests are reported: the Cragg-Donald F-statistic, which assumes homoskedastic i.i.d. errors, and the KleibergenPaap rk F-statistic, which is robust to firm-level clustering of standard errors. The Stock and Yogo (2005) critical value for instrument weakness is reported as well. If the Cragg-Donald F-statistic is larger than this critical value, one rejects the hypothesis that the actual maximal size of a 5% Wald test of joint significance of the endogenous regressors exceeds 10%. Critical values for the Kleibergen-Paap rk F-statistic as well as for individual regressors are not available in the literature to the best of my knowledge. The table shows that the model with all four instruments marginally rejects instrument weakness for the joint test, while the models with fewer instruments fail to reject. In sum, it appears fair to conclude that IV weakness-robust inference methods are necessary to draw conclusions. Fully robust inference can be achieved using the Anderson-Rubin (AR) 1949 statistic. This methodology is described in detail in Stock et al. (2002). In a nutshell, defining y as the dependent variable, Y as the endogenous regressors with coefficients β, and X1 as the exogenous 34 To justify the application of the linear model to the log transformation of financial covenant intensity, I re-estimate the results obtained thus far using ordinary least squares. The results are very similar qualitatively and quantitatively to those presented in the previous sections. 35 Note that this is not the same as plugging the predicted probabilities into the second stage, which would amount to a case of “forbidden regression” (Wooldridge, 2002). 36 Estimation uses the Stata module IVREG2 provided by Baum et al. (2010). 27 regressors and X2 as the instruments for Y , one can test the hypothesis β = β0 by running the regression: y − Y β0 = X1 γ1 + X2 γ2 + η, (5) and performing a Wald test of γ2 = 0 to obtain AR(β0 ), which is distributed FK,T −K . This test always has the correct size, regardless of instrument weakness. However, it loses power when instruments are weak, which makes it more difficult to find any significant effect.37 As shown in table 9, the AR-statistic strongly rejects the hypothesis that the relationship dummies jointly equal zero. To determine whether they are individually significant, the AR-statistic can be inverted to construct a fully robust confidence set. For example, the 95% confidence set contains all β0 for which AR(β0 ) fails to reject at the 5% significance level. For each of the models in table 9, I find the AR confidence set using a grid search that allows either relationship effect to range from [−10, 5] at increments of 0.05. In percentage terms, the true parameters are allowed to be located anywhere between -99.9% to +15,000%, a range that should include any reasonable value. Implementing a finer grid is simple, but will require additional computation time. Figure 3 shows the 95% and 90% AR confidence sets. Consistent with the limited power of the AR test under instrument weakness, the confidence sets are large and they become smaller as one moves from model 1 to models 2 and 3, where the instruments have been shown to be somewhat stronger. Importantly, however, the AR confidence sets are strictly contained in the third quadrant, with the exception of the 95% confidence set for model 3, which touches the zero axis for the coefficient on high relationship intensity. The figure also shows the highest p-values encountered at any grid point at or above zero. This p-value is always below 0.01 for low relationship intensity and it ranges between 0.004 and 0.057 for high relationship intensity. The upshot is that the IV estimation does not provide precise point estimates for the relationship effects, but even under weak-instrument robust inference, the effects are statistically significant. 37 In the presence of overidentifying restrictions, the AR-test assesses the joint hypothesis that β = β0 and that the overidentifying restrictions are valid. Because I use two predicted probabilities as instruments for two endogenous regressors, the AR-statistic reduces to a test of β = β0 . 28 6 Covenant tightness In this section, I explore the effects of relationship intensity on covenant tightness. As with covenant intensity, to the extent that lenders hold up relationship borrowers at the contracting point, covenant tightness should increase in relationship intensity. GaÌ‚rleanu and Zwiebel (2009) study a model of covenant tightness where tightness increases in borrower-lender information asymmetries and decreases in renegotiation costs. This generates conflicting predictions for lending relationships since it is likely that relationships decrease both information asymmetries and renegotiation costs. In terms of creating monitoring incentives, covenant tightness may be less important than covenant intensity. In the model of Rajan and Winton (1995), the monitoring incentive is created by giving the lender covenants that allow her to renegotiate loan contract terms and demand liquidation of poor projects if she monitors the borrower. Covenants need not be excessively tight, however. The crucial point is that the covenant is tight enough for the lender to gain control in bad states of nature. Previous work has shown that bank debt covenants are quite tight in general (Chava and Roberts, 2008). Hence monitoring incentive considerations may not create much variation in tightness across loans. In addition, it is unlikely that tightness is systematically affected by borrowers’ concern about covenant-created hold-up opportunities. This is due to two reasons. First, the violation of a particularly tight covenant is unlikely to lead outside lenders to a revision of their opinion about the borrower. The tighter the covenant, the more likely the violation arises from random variation. Second, Demiroglu and James (2010) show that covenants are set more tightly for firms with positive private information about their future prospects. Thus, tight covenants could in fact reduce hold-up problems from the perspective of the borrower since they signal the borrower’s quality to outside investors. I measure covenant tightness as follows. As discussed in Murfin (2010), for a covenant that stipulates a minimum value r for the financial ratio r that is normally distributed with standard deviation σ, tightness can be measured as the probability of a covenant violation: rt − r p=1−Φ , σ (6) where Φ denotes the cumulative standard normal distribution function. If the covenant limits r to a maximum ratio, the numerator in the parentheses (the covenant slack) is multiplied 29 by minus one. I use this equation to estimate the tightness of each financial covenant that is attached to a loan.38 Slack is measured using the difference between the financial ratio in the quarter prior to the loan start date and the covenant trigger. The quarterly standard deviation is estimated using the past twelve quarters.39 Exactly which covenants are set tightly may depend on the individual borrower’s situation and the purpose of the loan. For example, if the loan is used to finance a leveraged buy-out, the debt to EBITDA covenant is likely to be most important (Ivashina and Kovner, 2010). Therefore, I measure each loan’s covenant tightness as the tightness of the tightest covenant.40 As shown in the appendix (table A-1), this measure strongly predicts actual covenant violations over the course of the loan. Moreover, its predictive power is independent of covenant intensity, which also is a strong predictor on its own. This suggests that tightness and intensity are used to accomplish different goals and are not mere substitutes. Regressions of covenant tightness on relationship intensity are shown in table 10. Tightness is marginally negatively related to relationship intensity. No non-linear effect is found. This is consistent with tightness being driven by borrower-lender information asymmetries, which are reduced as the lender learns the borrower’s soft information. However, none of the interactions with various proxies for information asymmetries is statistically significant. There is reason to expect OLS estimates of the relationship effect on covenant tightness to be biased. As discussed earlier, Demiroglu and James (2010) show that covenants are set more tightly when borrowers and lenders have private information indicating future improvements in the firm’s financial performance, whereas covenant intensity does not exhibit this pattern. Private information in turn is likely to affect the choice to borrow from a relationship lender as 38 I exclude senior leverage covenants and senior debt to EBITDA covenants since quarterly data on senior debt is unavailable. I also exclude loan to value covenants, since they are based on valuations that are not available to me. This exclusion is unlikely to matter: senior leverage covenants are used in nine loans, and loan to value covenants are used in twelve loans (see table 1). Senior debt to EBITDA covenants are used in 11% of the loans, but they are almost always used in conjunction with a debt to EBITDA covenant, for which I do have data. 39 Data on intangible assets is frequently unavailable in Compustat Quarterly, which poses a problem in estimating the tightness of tangible net worth and debt to tangible net worth covenants. For cases where data on intangible assets is unavailable, I measure the tightness of these covenants using the covenant slack as of the end of the previous fiscal year and using the median standard deviation of the financial ratio for all firms in the borrower’s two-digit SIC industry whose book value of total assets are within a range of plus or minus 25% of the borrower’s total assets. 40 As mentioned by other researchers (e.g. Dichev and Skinner (2002), Chava and Roberts (2008), Murfin (2010)), some covenants appear to be violated at loan origination, which may be due to heterogeneity in covenant definitions and unmeasurable adjustments. I exclude such covenants from the analysis, and measure tightness using the data on the remaining covenants of the loan. This is likely to make the tightness measure more noisy. I deem this method preferable to excluding the entire loan, which might induce selection bias since loans with higher covenant intensity are more likely to have one covenant that appears violated. 30 well. I address this problem by estimating 2SLS regressions shown in table 11. The table does not show the first stage regressions as these are similar to the ones shown in the IV estimation for covenant intensity (table 9). The difference is that the first stage is linear here, since the endogenous variable I use is continuous. Using a relationship dummy with a first stage probit instead does not affect conclusions. Column (1) in table 11 uses the log of the borrower-lender distance as the instrumental variable. The relationship coefficient is much larger than in the OLS regression and is negatively significant. The instrument weakness test statistics strongly reject the hypothesis that the true size of a 5% significance test exceeds 10%. Column (2) adds the same state dummy and the industry size and industry age measures as instruments in the first stage. The relationship coefficient is smaller, but still significant at the 10% level. However, the combined set of instruments is weaker than the distance measure alone. The hypothesis that the instruments are weak is not rejected. The Hansen’s J test does not reject the validity of the overidentifying restrictions. Since the just-identified estimation is better behaved, columns (3) through (10) use the log of the borrower-lender distance as the instrument (see also Angrist and Pischke (2009) for a discussion of the median-unbiasedness of just-identified IV). Columns (3) through (10) show results of IV estimation for the interaction of relationship intensity with measures of information asymmetries and certification needs. As discussed in Wooldridge (2002), it is easy to construct an additional instrument for these interaction terms: one simply uses a linear regression to predict relationship intensity, then interacts the predicted values with the information asymmetry measure and uses this fitted interaction term as an instrument in the first stage of the 2SLS estimation. Columns (3) through (5) show that the reduction in covenant tightness is less strong for highly rated borrowers that have access to the commercial paper market, and stronger for small borrowers, consistent with tightness being driven by information asymmetries. These results are statistically significant at the 10% level. The coefficient on the interaction term for borrowers with a rating is positive, but not significant. Column (6) shows that the reduction in covenant tightness is concentrated in syndicated loans and is less pronounced in sole lender loans, an effect that is statistically significant at the 5% level. This is consistent with syndicated loans 31 having a greater need for certification by the lead arranger, where loan participants are more comfortable with looser covenants when the lead arranger knows the borrower well. Covenant definitions can vary across loans, which likely induces noise in the estimation. For this reason, columns (7) through (10) present estimates based on a tightness measure that only uses covenants for which the definitions are relatively standard across loans: net worth, tangible net worth, EBITDA, debt to EBITDA, and liquidity covenants.41 These are likely to be measured with less noise. As shown in columns (6) through (10), all results are stronger when focusing on these covenant types. The interaction terms for the rating dummy, access to the commercial paper market, small borrowers and sole lender loans are all statistically significant and the signs are consistent with the information asymmetry theory. I also test for nonlinearity in the relationship effect using the ordered probit IV strategy from table 9 (omitted from the table for brevity). I do not find evidence of such a nonlinear effect. As a further validity check of the IV estimation for interaction terms, I also repeat this estimation for the all-in spread drawn, to which the private information effect uncovered by Demiroglu and James (2010) does not apply (i.e. firms with positive private information about themselves are unlikely to accept higher yield spreads as a “signal” of their quality). For yield spreads, the IV estimation of interaction terms yields the same conclusions as OLS estimation. In sum, the results in this section suggest that the reduction in information asymmetries during a relationship allows for the setting of looser covenants. This is consistent with the private information content of covenant tightness found by Demiroglu and James (2010) and with the finding of Ivashina and Kovner (2010) who show that banking relationships result in a looser debt to EBITDA covenant for loans used to finance leveraged buy-outs.42 It also provides evidence supportive of the result in Demiroglu and James (2010) that covenant intensity and covenant tightness have different functions. 41 To determine this, I read the covenant definitions in the loan contracts collected by Nini et al. (2009). For each covenant type I randomly sampled twenty contracts containing the covenant from the intersection of my sample with their sample. I classify covenants as relatively standard if at least fifteen of the twenty covenants use the same definition (disregarding unmeasurable non-GAAP adjustments). For all covenant types that do not meet this criterion, less than ten covenants use the same definition, hence this seems a natural cut-off point. Also note that recent studies focus on the same covenant types as they are easier to measure (e.g. Chava and Roberts (2008) use net worth, tangible net worth, and current ratio covenants, and Demiroglu and James (2010) use current ratio and debt to EBITDA covenants). 42 The results presented here do not change when excluding leveraged buy-outs. This suggests that the importance of information asymmetries for covenant tightness holds in a more general setting. 32 7 Further robustness checks I perform a number of further robustness tests. First, the final sample of loans for which all necessary information is available is only half as large as the sample used to calculate relationship intensity. One may wonder whether there is something special about the loans that do not end up in the final sample. To this end, table 12 details why these loans are excluded. In 29% of the cases, covenant information is missing in DealScan. Since data availability for financial covenants requires the sample to start in 1995 and the relationship intensity measure uses a five-year lookback period, loans made in 1990-1994 are used to determine relationship intensity, but are not used in the final sample. These make up 23% of the excluded loans and are not a concern. Company financials and loan maturities are each missing in 14% of the cases. In 13% of the cases, relationship status is unknown as there was no loan in the previous five years. The other reasons are of negligible magnitudes. From this analysis, one might be concerned that there is something special about a) loans for which covenant information is missing and b) loans which are a borrower’s first loan in five years. I estimate Heckman sample selection models for equation 4 to address these concerns. Results are displayed in table 13. The first column shows results using a simple OLS estimation of equation 4 for comparison purposes since the Heckman sample selection procedure also uses a linear model as the outcome equation. These results are highly similar to the Poisson regression shown earlier. The second and third columns show the selection and outcome equations of the Heckman sample selection model that addresses sample selection according to the availability of covenant information. It seems unlikely that one could find a reasonable instrument for the availability of covenant information. According to DealScan officials, most of the covenant data is taken from loan contracts filed with the Securities Exchange Commission (SEC). The SEC regulation S-K requires material contracts to be filed as exhibits (§229.601). While measures of materiality can be conceived (e.g. the size of the loan relative to the size of the firm), it seems unreasonable to assume that the materiality of the loan is unrelated to covenant intensity. Thus, the Heckman selection model relies on the non-linearity in the first stage probit for identification. This may result in collinearity that could reduce the power of the model in finding a selection effect. The results for the outcome equation are nearly identical to those in the OLS regression. 33 The χ2 test does not reject the null hypothesis that the selection and outcome equations are independent. Hence, the fact that covenant information is missing for some loans does not appear to be an issue. The availability of covenant information is strongly predicted by the number of lenders involved in a loan, consistent with loans with many participants being relatively large and complex and thus more likely to be filed. Loan amounts, on the other hand, appear negatively related to the availability of covenant information. Note, however, that loan amounts, the number of lenders and the borrower’s size are all correlated. In any case, this does not affect the relationship results.43 Columns (4) through (6) examine the effect of omitting the loans for which relationship status is unknown because the loan is the borrower’s first loan recorded in DealScan in the past five years. In column (4), this is done by an OLS regression that includes a dummy variable that equals one for such loans. This dummy is negative significant, but the relationship terms retain roughly the same coefficients. Since we now know that first-time borrowers are different from the average, if the Heckman selection model has sufficient power despite the lack of an instrument, we would expect it to pick up this difference. Columns (5) and (6) show that it does. The χ2 test strongly rejects the hypothesis that the selection and outcome equations are independent. Nevertheless, after accounting for the selection effect, the relationship dummies are still negative significant. The fact that the Heckman selection model does pick up this difference without using an instrument also lends support to the validity of the results for sample selection on missing covenant information. I perform various additional robustness tests. First, one may be worried that the results are driven by the credit boom that occurred before the onset of the financial crisis in mid-2007. To test this, I split the sample into two parts, separating the observations during the credit boom from those outside of it. Following Kahle and Stulz (2010), I define the period from the year 2005 through mid-2007 as the credit boom period. I do not find evidence that the credit boom drives the results (regressions not reported). Allowing the credit boom to start in 2004 does not change these results. Many loans have a relationship intensity of either zero or one. To check whether this affects results, I re-estimate the regressions in table 4 for only those loans where relationship intensity 43 Replacing the loan amount with the materiality of the loan or dummy variables for various levels of materiality does not affect conclusions. 34 is larger than zero and smaller than one. Results using the relationship dummies are similar and results using the quadratic specification are stronger when doing so. A further worry may be that results are sensitive to the definition of financial covenant intensity. The results presented thus far use a simple count of the number of financial covenants attached to the loan. I also estimate the regressions in table 4 using a binary dependent variable that equals one if the loan contains more than two financial covenants (the sample median), and zero otherwise. An alternative way of counting covenants is to consider the six groups presented in table 1 and adding one for each covenant group included in the loan: debt to balance sheet, coverage, debt to cash flow, liquidity, net worth, and EBITDA covenants. This avoids potential double counting of similar covenants. Results are robust to these changes. A related concern is that some covenant types may trigger violations more frequently than others. For an extreme example, suppose that the only covenants that are ever violated are net worth, current ratio, and EBITDA covenants. In this case, covenant intensity results based on the simple count of all covenants would not be very informative. While I have argued and found evidence that a particular loan’s covenant tightness is affected differently by a lending relationship than covenant intensity, one can assess this issue using the average tightness of the various covenants across all loans. If some covenants are systematically more important than others, they will have a higher average tightness. It turns out that the vast majority of covenant types have an average violation probability of 10-15%. Among the covenants that are included in more than one percent of the loans, only the leverage covenant has a substantially smaller average violation probability of 6%. Results using an alternative covenant intensity measure that weights covenants by their average violation probability are highly similar to those using the simple count. Finally, loan contract terms such as covenant intensity, tightness, loan maturity, amount, collateral, syndicate size, and yield spreads are likely to be determined jointly. Estimating a system of equations for all these terms does not seem practical due to the lack of credible instruments. Following the suggestion of Murfin (2010), I estimate the relationship effects including all these variables as controls as well as excluding them. Results for covenant intensity and tightness are qualitatively and quantitatively highly similar regardless of whether these terms are included or excluded. 35 8 Conclusion In this paper, I study how banking relationships affect the structuring of financial covenants. Consistent with a decrease in monitoring and renegotiation costs and with covenants being used as an incentive to monitor, I find that financial covenant intensity increases when a borrower and a lender develop a lending relationship. However, I find that when a relationship becomes exclusive, concerns about hold-up opportunities created by covenant violations prevail and covenant use is reduced. This reduction in covenant use for exclusive relationships is contingent on the borrower’s ex ante bargaining power and is concentrated in large borrowers with access to the public debt market. Since covenants provide monitoring incentives by giving the lender control rights that competing claim holders do not have and cannot free-ride on, I argue that they should be a more effective means of incentivization if the lead arranger does not need to share the covenant benefits with other lenders who could free-ride on the lead arranger’s monitoring. Consistent with this, the increase in covenant intensity in a lending relationship is stronger for sole lender loans and loans with only one lead arranger. I also find some evidence that information acquisition by multiple lead arrangers and the resulting within-syndicate competition limits the lenders’ ability to hold up the borrower when a covenant breach occurs. The measured relationship effects are robust to various ways of accounting for endogeneity. In contrast to covenant intensity, I find that covenant tightness is driven by information asymmetries. Covenant tightness decreases in a lending relationship, especially if borrowers are opaque and if participant lenders have a need for certification by the lead arranger. These results are stronger after accounting for endogeneity, consistent with the finding of Demiroglu and James (2010) that covenant tightness, but not covenant intensity, contains private information about the borrower’s prospects, especially when information asymmetries are large. In sum, the results show that soft information acquisition changes the way in which hardinformation based monitoring tools are used. They also show that even large and rated borrowers worry about state-contingent hold-up by their lenders, although prior research suggests that they are not subject to hold-up when entering the loan contract. Finally, they show that covenant intensity and covenant tightness are both affected by the presence of soft information, but the difference in their usage means that they are affected very differently. 36 References Abadie, A. and G. W. Imbens (2006). Large sample properties of matching estimators for average treatment effects. Econometrica 74 (1), 235–267. Abadie, A. and G. W. Imbens (2008). On the failure of the bootstrap for matching estimators. Econometrica 76 (6), 1537–1557. Abadie, A. and G. W. Imbens (2009). Matching on the estimated propensity score. Technical report. Aghion, P. and P. Bolton (1992). An incomplete contracts approach to financial contracting. The Review of Economic Studies 59 (3), 473–494. Ai, C. and E. C. Norton (2003). Interaction terms in logit and probit models. Economics Letters 80 (1), 123–129. Anderson, T. W. and H. Rubin (1949). Estimation of the parameters of a single equation in a complete system of stochastic equations. The Annals of Mathematical Statistics 20 (1), 46–63. Angrist, J. D. and J.-S. Pischke (2009). Mostly Harmless Econometrics. Princeton, New Jersey: Princeton University Press. Baum, C. F., M. E. Schaffer, and S. Stillman (2010). ivreg2: Stata module for ex- tended instrumental variables/2SLS, GMM and AC/HAC, LIML and k-class regression. http://ideas.repec.org/c/boc/bocode/s425401.html. This version 3.0.05 (17 June 2010). Beneish, M. D. and E. Press (1995). The resolution of technical default. The Accounting Review 70, 337–353. Berger, A. N. and G. F. Udell (1995). Relationship lending and lines of credit in small firm finance. Journal of Business 68 (3), 351–381. Bharath, S. T., S. Dahiya, A. Saunders, and A. Srinivasan (2009). Lending relationships and loan contract terms. Review of Financial Studies, forthcoming. Boot, A. W. A. (2000). Relationship banking: What do we know? Intermediation 9 (1), 7–25. 37 Journal of Financial Carey, M. and M. Hrycay (1999). Credit flow, risk, and the role of private debt in capital structure. Working paper, Federal Reserve Board. Chava, S. and M. R. Roberts (2008). How does financing impact investment? The role of debt covenants. The Journal of Finance 63 (5), 2085–2121. Chen, K. C. W. and K. C. J. Wei (1993). Creditors’ decisions to waive violations of accountingbased debt covenants. The Accounting Review 68, 218–232. Dass, N. and M. Massa (2010). The impact of a strong bank-firm relationship on the borrowing firm. Review of Financial Studies, forthcoming. Degryse, H. and P. V. Cayseele (2000). Relationship lending within a bank-based system: Evidence from European small business data. Journal of Financial Intermediation 9 (1), 90–109. Demiroglu, C. and C. James (2010). The information content of bank loan covenants. Review of Financial Studies, forthcoming. Detragiache, E., P. Garella, and L. Guiso (2000). Multiple versus single banking relationships: theory and evidence. Journal of Finance 55 (3), 1133–1161. Dewatripont, M. and J. Tirole (1994). A theory of debt and equity: Diversity of securities and manager-shareholder congruence. The Quarterly Journal of Economics 109 (4), 1027–1054. Diamond, D. (1984). Financial intermediation and delegated monitoring. Review of Economic Studies 51 (3), 393–414. Diamond, D. (1991). Monitoring and reputation: The choice between bank loans and directly placed debt. Journal of Political Economy 99 (3), 689–721. Dichev, I. D. and D. J. Skinner (2002). Large-sample evidence on the debt covenant hypothesis. Journal of Accounting Research 40 (4), 1091–1123. Drucker, S. and M. Puri (2009). On loan sales, loan contracting, and lending relationships. Review of Financial Studies 22 (7), 2835–2872. Fama, E. F. (1985). What’s different about banks? Journal of Monetary Economics 15, 29–37. 38 GaÌ‚rleanu, N. and J. Zwiebel (2009). Design and renegotiation of debt covenants. Review of Financial Studies 22 (2), 749–781. Greenbaum, S. I., G. Kanatas, and I. Venezia (1989). Equilibrium loan pricing under the bank-client relationship. Journal of Banking and Finance 13 (2), 221–235. Guerin, C. A. (2007). The top 10 surprising things about syndicated loans. Real Estate Finance 24 (4), 13–15. Heckman, J. J., H. Ichimura, and P. E. Todd (1997). Matching as an econometric evaluation estimator: Evidence from evaluating a job training program. Review of Economic Studies 64 (4), 605–654. Heckman, J. J., H. Ichimura, and P. E. Todd (1998). Matching as an econometric evaluation estimator. Review of Economic Studies 65 (2), 261–294. Ioannidou, V. P. and S. Ongena (2010). ”Time for a change”: Loan conditions and bank behavior when firms switch banks. Journal of Finance, forthcoming. Ivashina, V. (2009). Asymmetric information effects on loan spreads. Journal of Financial Economics 92 (2), 300319. Ivashina, V. and A. Kovner (2010). The private equity advantage: leveraged buyout firms and relationship banking. Review of Financial Studies, forthcoming. Jensen, M. and W. Meckling (1976). Theory of the firm: Managerial behavior, agency costs, and ownership structure. Journal of Financial Economics 3 (4), 305–360. Kahan, M. and B. Tuckman (1993). Private vs. public lending: Evidence from covenants. Working paper no. 1152, UCLA Anderson Graduate School of Management. Kahle, K. M. and R. M. Stulz (2010). Financial policies and the financial crisis: How important was the systemic credit contraction for industrial corporations? Working paper 2010-03-013, Fisher College of Business. Lechner, M. (2001). Identification and estimation of causal effects of multiple treatments under the conditional independence assumption. In M. Lechner and F. Pfeiffer (Eds.), Econometric Evaluation of Labour Market Policies, pp. 43–58. Heidelberg: Physica. 39 Leuven, E. and B. Sianesi (2003). PSMATCH2: Stata module to perform full Mahalanobis and propensity score matching, common support graphing, and covariate imbalance testing. http://ideas.repec.org/c/boc/bocode/s432001.html. This version 4.0.4 (10 November 2010). Loughran, T. and J. Ritter (2004). Why has IPO underpricing changed over time? Financial Management 33 (3), 5–37. May, W. and M. Verde (2006). Loan volumes surge, covenants shrink in 2005. Technical report. Murfin, J. (2010). The supply-side determinants of loan contract strictness. Working paper, Duke University. Myers, S. C. (1977). Determinants of corporate borrowing. Journal of Financial Economics 5, 147–175. Nini, G., D. C. Smith, and A. Sufi (2009). Creditor control rights and firm investment policy. Journal of Financial Economics 92, 400–420. Nini, G., A. Sufi, and D. C. Smith (2010). Creditor control rights, corporate governance, and firm value. Working paper, University of Pennsylvania. Park, C. (2000). Monitoring and structure of debt contracts. The Journal of Finance 55 (5), 2157–2195. Petersen, M. A. and R. G. Rajan (1994). The benefits of lending relationships: Evidence from small business data. Journal of Finance 49 (1), 3–37. Rajan, R. G. (1992). Insiders and outsiders: The choice between informed and arm’s length debt. The Journal of Finance 50 (4), 1367–1400. Rajan, R. G. and A. Winton (1995). Covenants and collateral as incentives to monitor. The Journal of Finance 50 (4), 1113–1146. Rauh, J. D. and A. Sufi (2010). Capital structure and debt structure. Review of Financial Studies, forthcoming. Roberts, M. R. and A. Sufi (2009). Control rights and capital structure: An empirical investigation. The Journal of Finance 64 (4), 1657–1695. 40 Rosenbaum, P. R. and D. B. Rubin (1983). The central role of the propensity score in observational studies for causal effects. Biometrika 70 (1), 41–55. Schenone, C. (2010). Lending relationships and information rents: Do banks exploit their information advantages? Review of Financial Studies 23 (3), 1149–1199. Sharpe, S. A. (1990). Asymmetric information, bank lending and implicit contracts: A stylized model of customer relationships. Journal of Finance 45 (4), 1069–1087. Smith, C. W. (1993). A perspective on accounting-based debt covenant violations. The Accounting Review 68, 289–303. Smith, C. W. and J. B. Warner (1979). On financial contracting: An analysis of bond covenants. Journal of Financial Economics 7, 117–161. Sridhar, S. S. and R. P. Magee (1996). Financial contracts, opportunism, and disclosure management. Review of Accounting Studies 1, 225258. Stock, J. H., J. H. Wright, and M. Yogo (2002). A survey of weak instruments and weak identification in generalized method of moments. Journal of Business & Economic Statistics 20 (4), 518–529. Stock, J. H. and M. Yogo (2005). Testing for weak instruments in linear IV regression. In D. W. K. Andrews and J. H. Stock (Eds.), Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg, pp. 80108. Cambridge: Cambridge University Press. Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data. Cambridge, Massachusetts: The MIT Press. 41 Covenant intensity (incidence rate ratio) 1.02 1.04 1.06 1.08 1.1 1 0 20 40 60 Relationship intensity (%) Quadratic specification 80 100 Stepwise dummies Figure 1: Plot of the effect of relationship intensity on covenant intensity This figure plots the incidence rate ratio of financial covenant intensity for different levels of relationship intensity. The effect for the quadratic specification is plotted using the coefficients for the Relation (Max Amt) measure from table 4. The stepwise dummy specification shows the effect using ten dummy variables that equal one if relationship intensity is at least 0% (omitted from the regression), 10%, 20%, 30%, ..., 90%, and zero if it is below that dummy’s threshold. The dummy specification (not reported in table 4) controls for the same variables as the quadratic specification. 42 Covenant intensity (incidence rate ratio) 1.02 1.04 1.06 1.08 1 0 20 40 60 Relationship intensity (%) Unrated firms 80 100 Rated firms Figure 2: Plot of the effect of relationship intensity on covenant intensity for rated vs. unrated firms This figure plots the incidence rate ratio of financial covenant intensity for different levels of relationship intensity, comparing firms with an S&P rating to those without a rating. Control variables are the same as in table 4. 43 Low Relation −6 −5 −4 −3 −2 −1 0 1 2 Anderson−Rubin Confidence Sets from IV Model 2 Low Relation −10−9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 Anderson−Rubin Confidence Sets from IV Model 1 −10 −9 −8 −7 −6 −5 −4 −3 High Relation 90% Confidence Set −2 −1 0 1 2 −6 −5 95% Confidence Set −4 −3 −2 −1 High Relation 90% Confidence Set Largest p−values along or beyond the axes: Low Relation: 0.001 High Relation: 0.025 0 1 2 95% Confidence Set Largest p−values along or beyond the axes: Low Relation: 0.000 High Relation: 0.004 (a) (b) Low Relation −6 −5 −4 −3 −2 −1 0 1 2 Anderson−Rubin Confidence Sets from IV Model 3 −6 −5 −4 −3 −2 −1 High Relation 90% Confidence Set 0 1 2 95% Confidence Set Largest p−values along or beyond the axes: Low Relation: 0.001 High Relation: 0.057 (c) Figure 3: Anderson-Rubin Confidence Sets from Instrumental Variables Estimation This figure shows instrument weakness-robust Anderson-Rubin confidence sets for the coefficients of the relationship indicator variables based on the IV estimation shown in table 9. Confidence sets in figures (a), (b), and (c) are based on the specifications in the IV models 1, 2, and 3, respectively. The confidence sets are constructed by inverting the Anderson-Rubin (AR) statistic for joint significance of the endogenous variables (see Stock et al. (2002)). This process uses a grid search allowing the coefficients for the Low Relation and High Relation dummies to vary from [−10, 5] at intervals of 0.05. The AR statistic is calculated for each point on the grid, and the 95% (90%) confidence set encompasses all points where the p-value for the AR statistic exceeds 0.05 (0.10). The figures also show the largest p-value that was found at any point on the grid at or above the zero line for Low Relation and High Relation, respectively. Note that the scale in figure (a) differs from those in figures (b) and (c) in order to include the entire confidence set. 44 Table 1: Frequency of financial and non-financial covenant types This table shows the frequency of inclusion of the different covenant types reported in Dealscan for the sample of loans incurred by nonfinancial and non-utility US borrowers from 1995 to 2008 for which covenant information is available in Dealscan. Percent Financial Covenants Debt to Equity Covenant Debt to Tangible Net Worth Covenant Leverage Ratio Covenant Loan to Value Covenant Senior Leverage Covenant Any Debt to Balance Sheet Covenant 0.76 10.72 17.61 0.11 0.15 28.88 Cash Interest Coverage Covenant Debt Service Coverage Covenant Fixed Charge Coverage Covenant Interest Coverage Covenant Any Coverage Covenant 1.27 8.07 40.38 41.10 79.46 Debt to EBITDA Covenant Senior Debt to EBITDA Covenant Any Debt to Cash Flow Covenant 57.42 10.90 59.84 Current Ratio Covenant Quick Ratio Covenant Any Liquidity Covenant 11.25 2.33 13.48 Net Worth Covenant Tangible Net Worth Covenant Any Net Worth Covenant 22.81 20.02 42.82 EBITDA Covenant Non-Financial Covenants Asset Sales Sweep Equity Issuance Sweep Debt Issuance Sweep Any Sweep Provision 9.24 35.29 23.59 25.79 38.13 Capital Expenditure Restriction 22.33 Dividend Covenant 77.80 Observations 7923 45 Table 2: Comparison of the number of loans per firm used in the final sample and the number of loans used to determine relationship intensity This table compares the number of loans per firm that enter the final sample with the number of loans per firm that are available in the sample used to determine relationship intensity. Figures shown are for the firms that have at least one loan that enters the final sample. There is no firm with only one loan in the relationship sample since the definition of relationship intensity requires at least two loans (relationship intensity cannot be determined for the first loan agreement that a firm enters). Final sample Loans per firm Number Percent Relationship sample Number Percent 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 or more 1292 736 405 287 193 116 60 44 17 8 7 1 1 2 0 40.77 23.22 12.78 9.06 6.09 3.66 1.89 1.39 0.54 0.25 0.22 0.03 0.03 0.06 0.00 0 753 602 443 332 274 209 160 107 78 55 42 30 22 62 0.00 23.76 19.00 13.98 10.48 8.65 6.60 5.05 3.38 2.46 1.74 1.33 0.95 0.69 1.96 Total 3169 100.00 3169 100.00 46 Table 3: Univariate Tests of Differences in Firm Characteristics and Debt Issue Characteristics Conditional on Relationship Intensity This table presents averages and univariate t-tests of differences in covenant use, loan, and firm characteristics across three categories of relationship intensity. I classify relationship intensity as “low” if relationship intensity as measured by Relation (Max Amt) is less than 30%, “high” if it is at least 70%, and “medium” if relationship intensity is between 30% and 70%. Relation (Max Amt) and Relation (Sum Amt) denote the proportion of the total amount borrowed in the previous five years where the current lead arranger acted as a lead arranger. For previous loans with more than one lead arranger, Relation (Max Amt) gives each lead arranger credit for the full loan amount, and it equals the maximum of the lead arrangers’ relationship intensities. Relation (Sum Amt) gives each of N lead arrangers on a previous loan credit for 1/N times the amount of the loan and takes the sum of relationship intensities across lead arrangers. FinCov and NonFinCov count the number of financial and non-financial covenants included in the loan, respectively. Tightness denotes the tightness of the tightest financial covenant attached to the loan. For each covenant type, the probability of a covenant violation is estimated by evaluating the cumulative normal distribution function using the slack of the covenant in the quarter immediately prior to the start date of the deal divided by the standard deviation of the corresponding financial ratio over the previous twenty quarters. The weighted average maturity and yield spread over LIBOR for each dollar drawn down on the loan are given by Maturity and AllInDrawn, respectively. Collateral is a dummy variable that equals one if at least one of the facilities that form a loan is collateralized and zero otherwise. Loan amount is the total amount of the deal, and Assets are the borrowing firm’s total assets. All dollar amounts are converted to 2008 US dollars using the Consumer Price Index for all urban consumers. Leverage is defined as the book value of debt divided by total assets. Tangibility is the ratio of net property, plant and equipment to total assets. Rating is a categorical variable that equals zero if the firm has no S&P long-term issuer credit rating, 1, 2, 3, 4, if the rating is AAA, AA+, AA, AA-, respectively, and so on. Not rated is a dummy variable that equals one if the firm has no S&P rating and zero if it does have a rating. MTB is the market-to-book ratio, calculated as the market value of the firm’s shares outstanding plus the book value of debt and preferred stock divided by the book value of assets. Current Ratio is the ratio of current assets to current liabilities and Coverage Ratio is calculated as EBITDA divided by interest expense. Membership in the S&P 500 index is indicated by the dummy variable S&P 500. Low Relation (Max Amt) Relation (Sum Amt) FinCov Tightness NonFinCov AllInDrawn Maturity Collateral Loan amount Assets Leverage Tangibility Rating Not rated MTB Current Ratio Coverage Ratio S&P 500 Observations Medium High 0.0199 0.0169 2.5789 0.1755 1.9580 212.5960 45.0508 0.7137 337.5120 1538.5240 0.3122 0.3305 11.9204 0.6410 1.3603 2.0235 18.6451 0.0801 0.5194 0.4741 2.6751 0.1674 1.9916 185.0169 46.8235 0.6258 569.2621 3602.3168 0.3483 0.3496 11.4808 0.4549 1.4000 1.9321 13.7383 0.1635 0.9644 0.8887 2.4845 0.1426 1.7394 158.8755 45.5494 0.5476 630.6822 4255.0816 0.3017 0.3326 10.5923 0.4799 1.4615 1.8762 19.8726 0.1908 2833 954 4136 47 M–L ∗∗∗ H–M H–L ∗∗∗ 0.4995 0.4450 0.9446∗∗∗ ∗∗∗ ∗∗∗ 0.4572 0.4146 0.8718∗∗∗ ∗ ∗∗∗ 0.0962 −0.1905 −0.0944∗∗∗ ∗∗∗ −0.0081 −0.0248 −0.0329∗∗∗ ∗∗∗ 0.0336 −0.2523 −0.2186∗∗∗ ∗∗∗ ∗∗∗ −27.5791 −26.1414 −53.7205∗∗∗ 1.7727∗ −1.2741 0.4986 −0.0879∗∗∗ −0.0782∗∗∗ −0.1661∗∗∗ 231.7501∗∗∗ 61.4201 293.1702∗∗∗ ∗∗∗ 2063.7928 652.7649 2716.5577∗∗∗ 0.0361∗∗∗ −0.0466∗∗∗ −0.0105∗ 0.0191∗ −0.0170∗ 0.0021 −0.4396∗∗ −0.8885∗∗∗ −1.3281∗∗∗ −0.1861∗∗∗ 0.0250 −0.1611∗∗∗ 0.0397 0.0615 0.1012∗∗∗ −0.0914 −0.0559 −0.1473∗∗∗ ∗∗∗ ∗∗∗ −4.9068 6.1343 1.2275 0.1106∗∗∗ 0.0834∗∗∗ 0.0272∗ Table 4: The Effect of Relationship Intensity on Financial Covenant Use This table reports Poisson regressions of financial covenant intensity on relationship intensity and control variables for the sample of loans incurred by non-financial, non-public administration, non-utility US borrowers from 1995 - 2008. Regressions in columns 1 through 3 use Relation (Max Amt) as the measure of relationship intensity, and regressions in columns 4 through 6 use Relation (Sum Amt). The independent variables are defined in table 3. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are z statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Relation (Max Amt) (1) Relation (2) (3) Relation (Sum Amt) (4) (5) (6) 0.0407∗∗∗ 0.2275∗∗∗ (3.83) (4.38) 0.0416∗∗∗ 0.1981∗∗∗ (3.89) (4.13) −0.1858∗∗∗ (−3.57) −0.1563∗∗∗ (−3.27) (Relation)2 Low Relation −0.0732∗∗∗ (−5.40) −0.0647∗∗∗ (−5.37) High Relation −0.0332∗∗ (−2.40) −0.0276∗∗ (−2.31) Ln(Loan Amount) −0.0159∗ −0.0164∗ −0.0166∗ −0.0158∗ −0.0165∗ −0.0163∗ (−1.87) (−1.93) (−1.96) (−1.85) (−1.94) (−1.93) Ln(Maturity) 0.0134 (1.15) Ln(Lenders) 0.0426∗∗∗ 0.0414∗∗∗ 0.0421∗∗∗ 0.0428∗∗∗ 0.0414∗∗∗ 0.0423∗∗∗ (5.83) (5.68) (5.79) (5.86) (5.68) (5.81) Ln(Assets) Leverage Tangibility 0.0144 (1.25) 0.0145 (1.25) 0.0132 (1.13) 0.0145 (1.25) 0.0143 (1.23) −0.0241∗∗∗ −0.0247∗∗∗ −0.0242∗∗∗ −0.0236∗∗∗ −0.0248∗∗∗ −0.0245∗∗∗ (−3.24) (−3.33) (−3.26) (−3.17) (−3.34) (−3.30) 0.0460 (1.32) 0.0440 (1.26) 0.0446 (1.27) 0.0470 (1.34) 0.0433 (1.24) 0.0447 (1.28) −0.0667∗∗ −0.0675∗∗ −0.0674∗∗ −0.0672∗∗ −0.0672∗∗ −0.0674∗∗ (−2.49) (−2.53) (−2.53) (−2.52) (−2.52) (−2.53) Current Ratio 0.0095∗∗ (2.26) 0.0090∗∗ (2.17) 0.0091∗∗ (2.19) 0.0094∗∗ (2.25) 0.0092∗∗ (2.21) 0.0093∗∗ (2.22) Ln(1+Coverage Ratio) 0.0210∗∗∗ 0.0216∗∗∗ 0.0215∗∗∗ 0.0209∗∗∗ 0.0213∗∗∗ 0.0215∗∗∗ (3.27) (3.38) (3.36) (3.25) (3.33) (3.35) Rating 0.0278∗∗∗ 0.0273∗∗∗ 0.0275∗∗∗ 0.0279∗∗∗ 0.0276∗∗∗ 0.0277∗∗∗ (7.55) (7.42) (7.47) (7.57) (7.47) (7.55) Not rated 0.3748∗∗∗ 0.3714∗∗∗ 0.3725∗∗∗ 0.3756∗∗∗ 0.3737∗∗∗ 0.3751∗∗∗ (7.90) (7.85) (7.88) (7.92) (7.89) (7.94) MTB −0.0097∗ −0.0099∗ −0.0100∗ −0.0097∗ −0.0097∗ −0.0095∗ (−1.71) (−1.74) (−1.75) (−1.70) (−1.70) (−1.68) S&P 500 −0.1677∗∗∗ −0.1689∗∗∗ −0.1688∗∗∗ −0.1672∗∗∗ −0.1684∗∗∗ −0.1680∗∗∗ (−7.53) (−7.63) (−7.63) (−7.51) (−7.58) (−7.57) 48 Table 4: The Effect of Relationship Intensity on Financial Covenant Use — Continued Relation (Max Amt) (1) Constant (2) ∗∗∗ 0.6370 (6.81) Relation (Sum Amt) (3) 0.6332 (6.82) ∗∗∗ (4) ∗∗∗ 0.7059 (7.50) 0.6340 (6.79) (5) ∗∗∗ (6) ∗∗∗ 0.6337 (6.82) 0.6951∗∗∗ (7.41) Industry effects Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Observations 7923 7923 7923 7923 7923 7923 49 Table 5: The Effect of Relationship Intensity on Financial Covenant Use Depending on the Distribution of Bargaining Power and Syndicate Structure This table reports Poisson regressions for the sample of loans incurred by non-financial, non-utility US borrowers from 1995 - 2008 to assess the extent to which the use of financial covenants reflects the borrower’s bargaining power. The dependent variable is the number of financial covenants included in the loan. Columns 1 through 4 use Relation (Max Amt) as the relationship variable, while columns 5 through 8 use Relation (Sum Amt). CP Access indicates that the borrower has access to the commercial paper market (proxied by a short-term credit rating of A-2 or better). Sole Lender is a dummy variable that equals one if the loan is made by only one lender and zero if the loan is made by a syndicate of lenders. Multiple Lead equals one if there are at least two lead arrangers involved in the loan, and zero if there is only one lead arranger (regardless of the number of participants). Control variables are the same as in table 4. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are z statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Relationship variable: Relation (Max Amt) 50 (1) (2) ∗∗∗ (3) ∗∗∗ −0.0711 (−5.08) Relationship variable: Relation (Sum Amt) (4) ∗∗∗ −0.0567 (−3.89) (5) ∗∗∗ −0.0833 (−5.28) (6) ∗∗∗ Low Relation −0.0549 (−2.79) High Relation 0.0010 (0.05) Low Relation × Rated −0.0290 (−1.06) −0.0034 (−0.14) High Relation × Rated −0.0691∗∗ (−2.54) −0.0582∗∗ (−2.43) −0.0281∗∗ −0.0265∗ −0.0402∗∗ (−1.97) (−1.84) (−2.47) Low Relation × CP Access −0.0264 (−0.53) High Relation × CP Access −0.0854∗ (−1.77) −0.0568 (−3.18) 0.0012 (0.07) (7) ∗∗∗ −0.0641 (−5.12) (8) ∗∗∗ −0.0495 (−3.82) −0.0819∗∗∗ (−5.65) −0.0230∗ −0.0235∗ −0.0448∗∗∗ (−1.83) (−1.89) (−3.00) 0.0166 (0.37) −0.0492 (−1.16) Table 5: The Effect of Relationship Intensity on Financial Covenant Use Depending on the Distribution of Bargaining Power and Syndicate Structure — Continued Relationship variable: Relation (Max Amt) (1) (2) (3) (4) Relationship variable: Relation (Sum Amt) (5) (6) ∗ (7) (8) ∗∗ Low Relation × Sole Lender −0.0718 (−1.77) −0.0873 (−2.29) High Relation × Sole Lender −0.0368 (−0.84) −0.0507 (−1.21) Low Relation × Multiple Lead 0.0433 (1.41) 0.0575∗∗ (2.22) High Relation × Multiple Lead 0.0262 (0.92) 0.0544∗∗ (2.26) −0.1380∗∗∗ (−2.89) CP Access −0.1772∗∗∗ (−4.32) 51 −0.0398 (−0.97) Sole Lender −0.0271 (−0.70) −0.0532∗∗ (−2.49) −0.0376 (−1.41) Multiple Lead Control variables Yes Yes Yes Yes Yes Yes Yes Yes Industry effects Yes Yes Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Yes Yes Observations 7923 7923 7923 7923 7923 7923 7923 7923 Table 6: The Effect of Relationship Intensity on Financial Covenant Use Depending on the Borrower’s Information Opacity This table reports Poisson regressions for the sample of loans incurred by non-financial, nonutility US borrowers from 1995 - 2008 to assess the extent to which the use of financial covenants reflects the borrower’s information opacity. The dependent variable is the number of financial covenants included in the loan. Relation (Max Amt) is used as the relationship variable, but results using Relation (Sum Amt) are qualitatively and quantitatively similar. Small Borrower is a dummy variable indicating that the borrower’s size is below the median size of borrowers in the same year. Hightech is a dummy variable that equals one if the borrower is a member of a hightech industry as defined in Loughran and Ritter (2004). Low Analyst indicates that the number of analysts covering the borrower according to the I/B/E/S detail files is below the median number of analysts for borrowers in that year. High Forecast Disp. equals one if the dispersion of analyst forecasts for a firm’s earnings per share (EPS) is above the sample median in that year, and zero otherwise. Forecast dispersion for a firm is measured as the standard deviation of EPS forecasts divided by the absolute value of the mean EPS forecast in the I/B/E/S summary file. The dummy variable NASDAQ indicates that the firm’s stock traded on NASDAQ at the point of contracting the loan, according to CRSP. Control variables are the same as in table 4. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are z statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. (1) (2) (3) ∗∗∗ (4) ∗∗∗ −0.0643∗∗∗ −0.0260∗ −0.0306∗∗ −0.0613∗∗∗ −0.0631∗∗∗ −0.0473∗∗∗ (−3.59) (−1.74) (−2.07) (−3.50) (−3.17) (−2.67) 0.0660∗∗ (2.44) Low Relation × S&P 500 −0.0261 (−0.63) High Relation × S&P 500 −0.0597 (−1.63) Low Relation × Hightech −0.0316 (−0.82) High Relation × Hightech −0.0162 (−0.40) −0.0743 (−3.58) −0.0555∗∗∗ (−3.10) High Relation High Relation × Small Borrower −0.0594 (−3.24) (6) ∗∗∗ −0.0798 (−4.24) 0.0228 (0.82) −0.0690 (−4.75) (5) ∗∗∗ Low Relation Low Relation × Small Borrower −0.0690 (−4.77) ∗∗∗ −0.0089 (−0.33) Low Relation × Low Analyst 0.0575∗∗ (2.21) High Relation × Low Analyst Low Relation × High Forecast Disp. 0.0022 (0.07) High Relation × High Forecast Disp. 0.0268 (0.98) 52 Table 6: The Effect of Relationship Intensity on Financial Covenant Use Depending on the Borrower’s Information Opacity — Continued (1) (2) (3) (4) (5) (6) Low Relation × NASDAQ −0.0280 (−1.04) High Relation × NASDAQ 0.0355 (1.29) Small Borrower −0.0371 (−1.34) Hightech 0.0439 (1.20) −0.0526∗∗ (−2.12) Low Analyst −0.0567∗∗ (−2.23) High Forecast Disp. NASDAQ 0.0042 (0.16) Control variables Yes Yes Yes Yes Yes Yes Industry effects Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Observations 7923 7923 7923 7688 5972 7599 53 Table 7: The Relationship Effect on Yield Spreads and Non-financial Covenants This table reports the effect of relationship intensity on the all-in spread drawn and non-financial covenants for the sample of loans incurred by non-financial, non-utility US borrowers from 1995 - 2008. Regressions (1) through (7) are ordinary least squares regressions where the dependent variable is the all-in yield spread over LIBOR that the borrower pays on each dollar drawn down on the loan. Regressions (8) and (9) are Poisson regressions where the dependent variable is the number of non-financial covenants attached to the loan. The regressions use the Relation (Max Amt) measure described in table 3 as the measure of relationship intensity. Control variables are the same as in table 4. Variables used in the interaction terms are explained in tables 5 and 6. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are t-statistics (regressions (1) through (7)) and z statistics (regressions (8) and (9)) adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ ∗∗ , , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. All-In Spread Drawn (1) Relation (2) −9.5481∗∗∗ (−3.64) 54 1.7109 (0.50) High Relation −6.7575∗∗ (−2.06) Relation × Small Borrower Relation × NASDAQ Relation × High Forecast Disp. Relation × Multiple Lead (4) −15.3146∗∗∗ −3.6334 (−4.22) (−1.06) Low Relation Relation × Rated (3) (5) −5.7208∗ (−1.82) Non-financial Covenants (6) 0.6450 (0.20) (7) −9.8641∗∗∗ (−3.27) (8) (9) −0.0547∗∗∗ (−2.78) −0.0125 (−0.51) −0.0590∗∗ (−2.37) 14.1106∗∗∗ (2.84) −10.2848∗∗ (−2.06) −3.2219 (−0.62) −8.8281∗ (−1.70) 0.7034 (0.12) Table 7: The Relationship Effect on Yield Spreads and Non-financial Covenants — Continued All-In Spread Drawn (1) (2) (3) (4) (5) Non-financial Covenants (6) (7) (8) (9) ∗∗∗ Small Borrower 22.5165 (4.32) 11.4741∗∗ (2.43) NASDAQ 26.7162∗∗∗ (6.25) High Forecast Disp. Multiple Lead 7.0896 (1.32) 55 Control variables Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 7649 7649 7649 7649 7371 5816 7649 7923 7923 Table 8: Propensity score matching This table presents propensity score matching estimates for the effect of relationship intensity on covenant intensity. Propensity scores are obtained from an ordered probit regression that determines the probability of whether a borrower’s relationship intensity will be low, medium, or high. Control variables in the ordered probit regression include borrower characteristics as in table 4, the log of the loan amount, as well as loan purpose, loan type, loan year and industry fixed effects. The resulting propensity scores are used to match loans in the treatment group with loans in the untreated group. The Nearest neighbor estimators calculate the difference in financial covenant intensity between each treated loan and the n loans in the untreated group with the closest propensity scores. The Gaussian and Epanechnikov estimators perform kernel matching using a Gaussian and Epanechnikov kernel, respectively. Both kernels assign more weight to untreated loans whose propensity scores are closer to the propensity score of the treated loan. The Gaussian kernel uses the entire set of untreated loans; the smaller the bandwidth, the faster the weights decline in the propensity score distance. The Epanechnikov kernel only uses untreated loans whose propensity score differs from the propensity score of the treated loan by no more than the specified bandwidth. I use a bandwidth of 0.01 for both kernels. In columns (1) through (3), the treated group are loans with medium relationship intensity and the untreated group are loans with low relationship intensity. In columns (4) through (6), loans with high relationship intensity are defined as the treated group and loans with medium relationship intensity are defined as the untreated group. In column (1), loans of medium relationship intensity are matched to all loans of low relationship intensity. In column (2), the untreated group is restricted to loans where the previous loan was a relationship loan, but the current loan is not, thus “breaking” the relationship. In column (3), the untreated group is restricted to loans of low relationship intensity that do not break an existing relationship. Column (4) uses all loans with high relationship intensity as the treated group, while column (5) uses loans where the borrower always borrowed from the same lender in the past five years as the treated group and column (6) uses loans with high relationship intensity, but where the borrower obtained a loan from a different lender at least once in the past five years. Standard errors for the nearest neighbor estimators are calculated using the Abadie-Imbens 2006 variance estimator. Standard errors for the kernel estimators are obtained by bootstrapping with 1,000 replications. Numbers in parentheses are t-statistics. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Medium Relation (1) (2) (3) High Relation (4) (5) (6) Nearest neighbor (n=10) 0.1415∗∗∗ 0.2155∗∗∗ 0.1412∗∗∗ −0.0833∗∗ −0.0805∗∗ −0.0921∗∗ (3.62) (4.57) (3.12) (−2.36) (−2.22) (−2.04) Nearest neighbor (n=50) 0.1433∗∗∗ 0.2034∗∗∗ 0.1424∗∗∗ −0.1024∗∗∗ −0.1010∗∗∗ −0.1069∗∗ (3.73) (4.35) (3.16) (−2.90) (−2.77) (−2.44) Gaussian 0.1127∗∗∗ 0.1861∗∗∗ 0.1002∗∗ −0.0879∗∗ −0.0877∗∗ −0.0885∗∗ (2.99) (3.86) (2.19) (−2.53) (−2.37) (−2.04) Epanechnikov 0.1327∗∗∗ 0.2037∗∗∗ 0.1302∗∗∗ −0.0797∗∗ −0.0793∗∗ −0.0809∗ (3.46) (4.45) (2.98) (−2.26) (−2.19) (−1.75) Treated obs. Untreated obs. 949 2833 949 984 949 1132 56 4141 949 3122 949 1019 949 Table 9: Instrumental Variables Estimation This table shows results for the relationship effects on financial covenant intensity from two-stage least squares (2SLS) estimation. Three different models are presented that use different sets of instruments. For each model, the column denoted 1st Stage shows the results of an ordered probit model, where the dependent variable equals one for low, two for medium, and three for high relationship intensity. Predicted probabilities for low and high relationship intensity from this model are used as instruments in the first stage of the 2SLS estimation. The column denoted IV shows the 2SLS result for financial covenant intensity. Relationship intensity is calculated using the Relation (Max Amt) measure defined in table 3. The instruments include the log of one plus the distance between the borrower’s headquarters and the headquarters of the closest lead arranger, a dummy variable that indicates whether the borrower’s and the lead arranger’s headquarters are in the same state, the log of the total assets of the median firm in the borrower’s Fama-French 38 industry in the year prior to entering the loan agreement, as well as the log of the age of the oldest firm in the borrower’s Fama-French 38 industry. Firm age is calculated as 2008 minus the year of the initial public offering as recorded in Compustat. Control variables are described in table 4. Cut 1 and Cut 2 refer to the cut points for the latent variable in the ordered probit model. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are t statistics (z statistics in the case of the ordered probit) adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. The Cragg-Donald F-statistic and the Kleibergen-Paap rk F-statistic are tests for instrument weakness, where the former assumes homoskedastic i.i.d. errors, while the latter is robust to clustered standard errors. The critical value from Stock and Yogo (2005) is the value that the Cragg-Donald F-statistic needs to exceed to reject the hypothesis that a test for the joint significance of the endogenous regressors with nominal size 5% has a true size of more than 10%. The Anderson-Rubin F-statistic tests whether the endogenous regressors are jointly zero and is robust to instrument weakness. Model 1 1st Stage Model 2 1st Stage IV Model 3 IV 1st Stage IV Low Relation −2.0129∗∗∗ (−2.87) −2.1060∗∗∗ (−3.36) −1.5934∗∗∗ (−3.06) High Relation −1.5611∗∗ (−2.07) −1.6120∗∗ (−2.51) −1.0471∗∗ (−2.02) Ln(1+Distance) Same State −0.0330∗∗ (−2.20) −0.0659∗∗∗ (−6.15) 0.2028∗∗∗ (3.22) −0.0326∗∗ (−2.18) 0.2041∗∗∗ (3.24) −0.0605∗∗ (−2.35) Ln(Industry Median Size) −0.0823∗∗∗ (−3.07) −0.1790∗∗ (−2.52) Ln(Industry Age) Ln(Loan Amount) 0.2260∗∗∗ −0.0559∗∗∗ 0.2210∗∗∗ −0.0584∗∗∗ 0.2243∗∗∗ −0.0537∗∗∗ (8.61) (−2.91) (8.43) (−2.86) (8.58) (−3.21) Ln(Assets) 0.0016 (0.06) Leverage 0.2241∗ −0.0197 (1.79) (−0.25) −0.0262∗ (−1.84) 57 0.0044 (0.16) −0.0259∗ (−1.76) 0.2219∗ −0.0252 (1.77) (−0.32) 0.0086 (0.31) −0.0256∗∗ (−2.05) 0.2199∗ −0.0048 (1.76) (−0.07) Tangibility Current Ratio Ln(1+Coverage Ratio) 0.0136 (0.14) −0.0153 (−1.00) −0.0822∗ (−1.69) 0.0052 (0.55) 0.0444 (0.46) −0.0127 (−0.83) 0.0798∗∗∗ 0.0348∗∗ (3.35) (2.35) −0.0829∗ (−1.67) 0.0048 (0.52) 0.0288 (0.30) −0.0148 (−0.96) 0.0786∗∗∗ 0.0343∗∗ (3.29) (2.35) −0.0783∗ (−1.84) 0.0084 (1.08) 0.0788∗∗∗ 0.0290∗∗ (3.30) (2.39) Rating −0.0340∗∗∗ 0.0204∗∗∗ −0.0335∗∗∗ 0.0205∗∗∗ −0.0344∗∗∗ 0.0234∗∗∗ (−3.03) (2.58) (−2.98) (2.68) (−3.07) (3.66) Not rated −0.4148∗∗∗ 0.3309∗∗∗ −0.4018∗∗∗ 0.3329∗∗∗ −0.4147∗∗∗ 0.3566∗∗∗ (−2.81) (3.78) (−2.72) (3.85) (−2.81) (4.87) MTB −0.0012 (−0.05) S&P 500 −0.1618∗∗ −0.2148∗∗∗ −0.1608∗∗ −0.2157∗∗∗ −0.1740∗∗ −0.1964∗∗∗ (−2.30) (−4.73) (−2.27) (−4.94) (−2.47) (−5.45) −0.0179 (−1.56) −0.0050 (−0.22) −0.0180 (−1.52) −0.0053 (−0.24) −0.0175∗ (−1.79) Ln(Maturity) 0.0553∗∗ (2.24) 0.0582∗∗ (2.21) 0.0555∗∗ (2.58) Ln(Lenders) 0.0118 (0.53) 0.0079 (0.31) 0.0049 (0.26) Constant 2.2062∗∗∗ (3.11) 2.2716∗∗∗ (3.60) 1.7901∗∗∗ (3.49) Cut 1 0.6248 (1.38) 0.6302 (1.33) 0.4041 (0.78) Cut 2 0.9784∗∗ (2.16) 0.9835∗∗ (2.07) 0.7583 (1.47) Industry effects Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Observations 6643 6643 6643 6643 6643 6643 Weak ID tests Cragg-Donald F-stat. Kleibergen-Paap rk F-stat. Stock-Yogo (2005) crit. 4.32 4.45 7.03 6.26 6.71 7.03 7.00 7.26 7.03 Weak ID robust inference Anderson-Rubin F-stat. Anderson-Rubin p-value 18.100 0.000 18.028 0.000 17.092 0.000 58 Table 10: The Effect of Relationship Intensity on Covenant Tightness This table shows OLS regressions of covenant tightness on relationship intensity and control variables. Covenant tightness is estimated as follows. For each covenant type, the probability of a covenant violation is estimated by evaluating the cumulative normal distribution function using the slack of the covenant in the quarter immediately prior to the start date of the deal divided by the standard deviation of the corresponding financial ratio over the previous twenty quarters. Each loan’s tightness is given as the tightness of the loan’s tightest covenant. Since information on intangible assets is frequently missing in Compustat Quarterly, I substitute the information for tangible net worth covenants and debt to tangible net worth covenants with the annual slack divided by the median standard deviation of the financial ratio for comparable firms with quarterly data. Comparable firms are defined as firms in the same two-digit SIC industry with total assets that differ by no more than plus or minus 25% from the total assets of the borrower. Relationship intensity is measured as Relation (Max Amt) defined in equation 1. Control variables are the same as in table 4. Variables used in the interaction terms are explained in tables 5 and 6. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are t-statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. (1) Relation (2) (3) −0.0090∗ (−1.83) −0.0083 (−1.29) Low Relation −0.0001 (−0.02) High Relation −0.0081 (−1.29) (4) (5) −0.0087∗ −0.0062 (−1.68) (−0.89) (6) −0.0113∗∗ (−2.07) −0.0019 (−0.20) Relation × Rated −0.0048 (−0.33) Relation × CP Access −0.0047 (−0.50) Relation × Small Borrower Relation × Sole Lender 0.0085 (0.69) −0.0089 (−0.59) CP Access 0.0203∗∗ (2.27) Small Borrower −0.0314∗∗∗ (−3.27) Sole Lender Ln(Loan Amount) −0.0007 (−0.19) Ln(Maturity) −0.0187∗∗∗ −0.0186∗∗∗ −0.0187∗∗∗ −0.0189∗∗∗ −0.0186∗∗∗ −0.0192∗∗∗ (−3.76) (−3.74) (−3.76) (−3.78) (−3.75) (−3.84) Ln(Lenders) 0.0017 (0.54) −0.0008 (−0.21) 0.0016 (0.51) 59 −0.0007 (−0.19) 0.0017 (0.54) −0.0007 (−0.18) 0.0016 (0.49) −0.0010 (−0.27) 0.0021 (0.67) −0.0019 (−0.52) −0.0049 (−1.35) Table 10: The Effect of Relationship Intensity on Covenant Tightness — Continued (1) Ln(Assets) Leverage (2) (3) (4) (5) −0.0101∗∗∗ −0.0101∗∗∗ −0.0101∗∗∗ −0.0101∗∗∗ −0.0058 (−3.00) (−3.01) (−3.00) (−3.01) (−1.54) (6) −0.0096∗∗∗ (−2.88) 0.0617∗∗∗ 0.0616∗∗∗ 0.0617∗∗∗ 0.0623∗∗∗ 0.0606∗∗∗ 0.0605∗∗∗ (3.72) (3.71) (3.72) (3.75) (3.65) (3.65) Tangibility −0.0626∗∗∗ −0.0625∗∗∗ −0.0626∗∗∗ −0.0624∗∗∗ −0.0626∗∗∗ −0.0635∗∗∗ (−5.55) (−5.55) (−5.55) (−5.54) (−5.57) (−5.64) Current Ratio −0.0109∗∗∗ −0.0109∗∗∗ −0.0109∗∗∗ −0.0110∗∗∗ −0.0109∗∗∗ −0.0109∗∗∗ (−5.81) (−5.83) (−5.81) (−5.87) (−5.79) (−5.81) Ln(1+Coverage Ratio) −0.0206∗∗∗ −0.0206∗∗∗ −0.0206∗∗∗ −0.0206∗∗∗ −0.0206∗∗∗ −0.0209∗∗∗ (−6.56) (−6.56) (−6.57) (−6.57) (−6.56) (−6.66) Rating 0.0107∗∗∗ 0.0107∗∗∗ 0.0107∗∗∗ 0.0099∗∗∗ 0.0106∗∗∗ 0.0104∗∗∗ (7.63) (7.62) (7.58) (6.37) (7.52) (7.41) Not rated 0.1433∗∗∗ 0.1432∗∗∗ 0.1418∗∗∗ 0.1337∗∗∗ 0.1392∗∗∗ 0.1384∗∗∗ (8.40) (8.40) (7.69) (7.03) (8.12) (8.11) MTB −0.0016 (−0.58) −0.0016 (−0.58) −0.0016 (−0.58) −0.0016 (−0.58) −0.0014 (−0.52) −0.0014 (−0.53) S&P 500 −0.0095 (−1.14) −0.0096 (−1.14) −0.0095 (−1.13) −0.0064 (−0.73) −0.0119 (−1.40) −0.0079 (−0.94) Constant 0.1327∗∗ (2.43) 0.1322∗∗ (2.41) 0.1337∗∗ (2.43) 0.1433∗∗∗ 0.1011∗ (2.62) (1.83) 0.1574∗∗∗ (2.81) Industry effects Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Observations Adj. R-squared 6947 0.144 6947 0.144 6947 0.144 6947 0.144 6947 0.145 6947 0.146 60 Table 11: Instrumental Variables Regressions for the Effect of Relationship Intensity on Covenant Tightness This table shows 2SLS estimates of the effect of relationship intensity on covenant tightness. Relationship intensity is measured as Relation (Max Amt) defined in equation 1. Columns (1) through (6) show results using the tightness measure as defined in table 10. Columns (7) through (10) measure tightness using only those covenants for which definitions across loans are relatively standardized (i.e. out of a random sample of twenty covenants, at least fifteen use the same definition (disregarding non-GAAP adjustments)): net worth, tangible net worth, EBITDA, debt to EBITDA, and liquidity covenants. For all other financial covenant types, less than ten contracts use the same definition. Column (1) uses the log of one plus the distance between the borrower and the nearest lead arranger as the instrument. Column (2) shows the result when adding the three other instruments: a dummy variable that indicates whether the borrower’s and the lead arranger’s headquarters are in the same state, the log of the total assets of the median firm in the borrower’s Fama-French 38 industry in the year prior to entering the loan agreement, as well as the log of the age of the oldest firm in the borrower’s Fama-French 38 industry. The estimations in columns (3) through (10) again use the log of the borrower-lender distance as the instrument for relationship intensity. The interaction terms are instrumented by regressing relationship intensity on the log of the borrower-lender distance and the control variables, and interacting the predicted value for relationship intensity with the interaction variable. The resulting variable is used as an instrument in the first stage of the 2SLS estimation, as discussed in Wooldridge (2002). Control variables are the same as in table 4. Variables used in the interaction terms are explained in tables 5 and 6. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are t-statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. 61 Tightness (1) Relation Relation × Rated Relation × CP Access (2) −0.1560∗∗ −0.0800 (−2.17) (−1.60) (3) (4) (5) −0.1694∗∗ −0.1615∗∗ −0.1222 (−2.24) (−2.23) (−1.57) (6) (7) (8) (9) (10) −0.2078∗∗∗ −0.2991∗∗∗ −0.2767∗∗∗ −0.2231∗∗ −0.3330∗∗∗ (−2.62) (−3.30) (−3.16) (−2.38) (−3.35) 0.0949∗∗ (2.15) 0.0353 (0.96) 0.0957∗ (1.73) 0.1774∗∗∗ (3.45) −0.0784∗ (−1.77) −0.0619 (−1.64) Relation × Small Borrower 0.1506∗∗ (2.28) Relation × Sole Lender CP Access Tightness (Narrow Def.) −0.0850∗ (−1.91) 0.1725∗∗ (2.26) −0.1450∗∗∗ (−3.60) Table 11: Instrumental Variables Regressions for the Effect of Relationship Intensity on Covenant Tightness — Continued Tightness (1) (2) (3) (4) Tightness (Narrow Def.) (5) (6) (7) (8) 0.0446∗ (1.77) Small Borrower (9) 0.0363 (1.17) −0.1024∗∗∗ (−3.32) Sole Lender (10) −0.1109∗∗∗ (−3.01) 62 Control variables Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Industry effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 5845 5845 5845 5845 5845 5845 4682 4682 4682 4682 18.66 14.71 7.03 18.60 14.78 7.03 18.65 14.88 7.03 17.71 13.96 7.03 15.03 11.75 7.03 14.80 11.66 7.03 14.86 11.80 7.03 13.94 10.95 7.03 Hansen’s J Hansen’s J p-value Weak ID tests Cragg-Donald F-stat. Kleibergen-Paap rk F-stat. Stock-Yogo (2005) crit. 2.76 0.431 37.30 29.66 16.38 17.24 13.50 24.58 Table 12: Reasons for exclusions of relationship sample loans from the final sample For all borrowers with at least one loan in the final sample, this table details why a loan that was used to determine relationship intensity was not included in the final sample. Since data availability for financial covenants requires the sample to start in 1995 and the relationship intensity measure uses a five-year lookback period, loans made in 1990-1994 are used to determine relationship intensity, but are not used in the final sample. For the same reason, loans where there was no previous loan the last five years are excluded from the final sample. Observations with a coverage ratio of less than minus one are excluded from the final sample since the log of one plus the coverage ratio is one of the control variables in the regressions. A few observations are excluded from the final sample because the firms are classified as financial, public administration or utility firms for these observations, even though they are classified as not belonging to these industries for other observations. Number Percent Covenant information missing Loan prior to 1995 Company financials missing Maturity missing No loan in previous five years No record in Compustat for fiscal year prior to loan start Coverage ratio less than minus one Link to Compustat unavailable for this package Foreign currency Firm classified as financial, public administration or utility during this year Syndication country foreign or unknown SIC code missing 2316 1815 1131 1086 1059 169 147 63 41 39 18 4 29.36 23.01 14.34 13.77 13.43 2.14 1.86 0.80 0.52 0.49 0.23 0.05 Total 7888 100.00 63 Table 13: Heckman selection models This table shows estimates of the effect of relationship intensity on financial covenant intensity using OLS regressions and Heckman selection models. The dependent variable is the log of the financial covenant count. Columns 1 through 3 ask whether a selection effect is present in the loans where information on financial covenants is missing. Columns 4 through 6 assess selection effects in loans that are excluded from the final sample because they constitute a borrower’s first borrowing (during the previous five years) in DealScan. The total number of observations in columns 4 and 5 does not match because some first borrowings also lack information on covenant intensity. Moving these observations to the analysis on missing covenant information yields qualitatively and quantitatively similar results. Control variables are the same as in table 4. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are z statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Missing covenant information OLS Heckman model Outcome Selection First borrowing OLS Heckman model Outcome Selection Low Relation −0.0822∗∗∗ −0.0822∗∗∗ (−5.32) (−5.34) −0.0818∗∗∗ −0.0631∗∗∗ (−5.29) (−4.66) High Relation −0.0368∗∗ −0.0368∗∗ (−2.35) (−2.36) −0.0359∗∗ −0.0274∗∗ (−2.29) (−1.99) −0.0421∗∗ (−2.08) First Borrowing Ln(Loan Amount) −0.0238∗∗ −0.0236∗∗ −0.0897∗∗∗ −0.0193∗∗ −0.0194∗ (−2.49) (−2.46) (−3.06) (−2.12) (−1.90) Ln(Maturity) 0.0127 (0.97) Ln(Lenders) 0.0557∗∗∗ 0.0544∗∗∗ 0.4180∗∗∗ 0.0521∗∗∗ (6.88) (6.46) (17.04) (6.68) Ln(Assets) Leverage Tangibility 0.0124 (0.96) 0.0383 (1.03) 0.0127 (1.05) 0.0092 (0.32) 0.0244∗ −0.0890∗∗ (1.79) (−2.38) 0.0245∗∗∗ 0.0845∗∗∗ (2.84) (3.20) −0.0230∗∗∗ −0.0223∗∗∗ −0.2071∗∗∗ −0.0236∗∗∗ −0.0338∗∗∗ 0.0728∗∗∗ (−2.69) (−2.61) (−7.42) (−2.85) (−3.73) (2.91) 0.0633 (1.59) 0.0642 (1.63) −0.0605∗∗ −0.0606∗∗ (−2.02) (−2.04) Current Ratio 0.0107∗∗ (2.18) Ln(1+Coverage Ratio) 0.0465 (1.22) 0.0183 (0.20) −0.0549∗ (−1.94) 0.0451 (0.39) −0.0613∗∗ −0.0074 (−2.04) (−0.10) 0.0101∗ −0.0038 (1.93) (−0.26) 0.0358∗∗∗ 0.0357∗∗∗ 0.0408 (4.75) (4.75) (1.63) 0.0345∗∗∗ (4.90) 0.0372∗∗∗ −0.0840∗∗∗ (4.78) (−3.93) Rating 0.0272∗∗∗ 0.0271∗∗∗ 0.0121 (6.63) (6.65) (1.11) 0.0283∗∗∗ (6.92) 0.0259∗∗∗ 0.0201 (6.14) (1.32) Not rated 0.3756∗∗∗ 0.3752∗∗∗ 0.1319 (7.18) (7.21) (0.93) 0.3863∗∗∗ (7.47) 0.3668∗∗∗ 0.1842 (6.82) (0.96) −0.0102 (−1.60) −0.0101 (−1.59) 64 0.0382∗∗ (2.16) 0.0475 (1.17) 0.0097∗∗ (2.10) MTB 0.0106∗∗ (2.16) −0.3015∗∗ (−2.44) −0.0220 (−0.92) −0.0108∗ (−1.78) −0.0076 (−1.12) −0.0206 (−1.11) Table 13: Heckman selection models — Continued Missing covenant information OLS Heckman model Outcome Selection S&P 500 −0.1725∗∗∗ −0.1727∗∗∗ 0.0498 (−7.25) (−7.29) (0.75) Constant 0.6180∗∗∗ 0.6239∗∗∗ 0.3962 (5.51) (5.55) (1.09) First borrowing OLS Heckman model Outcome Selection −0.1698∗∗∗ −0.1504∗∗∗ −0.0891 (−7.34) (−6.21) (−1.12) 0.6320∗∗∗ (5.83) 0.8542∗∗∗ 0.3659 (7.08) (1.02) Industry effects Yes Yes Yes Yes Yes Yes Year effects Yes Yes Yes Yes Yes Yes Loan purpose effects Yes Yes Yes Yes Yes Yes Loan type effects Yes Yes Yes Yes Yes Yes Observations Censored Observations Rho χ2 p-value 7923 10237 2314 −0.024 0.18 0.67 8708 8982 1059 −0.955 1023.51 0.00 65 Table A-1: Covenant intensity and tightness as predictors of covenant violations This table shows marginal effects from Probit regressions predicting whether a new covenant violation occurs during the time a loan is active. Quarterly covenant violation data are from Nini et al. (2010). As suggested in their paper, covenant violations are judged as new if the firm has not reported a violation in the previous four quarters. Ln(FinCov) is the log of the number of financial covenants attached to the loan (covenant intensity). Covenant tightness is defined as in table 10. Control variables are the same as in table 4. All regressions control for industry fixed effects at the one-digit SIC level, year fixed effects at the respective loan’s origination date, as well as loan purpose and loan type fixed effects. Numbers in parentheses are z statistics adjusted for heteroskedasticity and firm-level clustering. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels, respectively. (1) Ln(FinCov) 0.0598∗∗∗ (4.20) −0.0174∗ (−1.87) (3) 0.0588∗∗∗ (3.58) 0.1869∗∗∗ (5.28) Tightness Ln(Loan Amount) (2) −0.0183∗ (−1.84) 0.1665∗∗∗ (4.65) −0.0178∗ (−1.78) Ln(Maturity) 0.1153∗∗∗ (8.17) 0.1177∗∗∗ (7.88) 0.1164∗∗∗ (7.82) Ln(Lenders) 0.0029 (0.34) 0.0089 (0.97) 0.0064 (0.70) Ln(Assets) −0.0253∗∗∗ (−2.85) −0.0255∗∗∗ (−2.67) −0.0240∗∗ (−2.50) Leverage −0.0111 (−0.27) −0.0293 (−0.68) −0.0318 (−0.74) Tangibility −0.0015 (−0.04) −0.0041 (−0.12) −0.0008 (−0.02) Current Ratio Ln(1+Coverage Ratio) 0.0085 (1.53) −0.0130∗ (−1.69) 0.0116∗∗ (1.99) −0.0082 (−0.98) 0.0108∗ (1.85) −0.0095 (−1.14) Rating 0.0166∗∗∗ (3.89) 0.0156∗∗∗ (3.40) 0.0142∗∗∗ (3.05) Not rated 0.2300∗∗∗ (4.12) 0.2130∗∗∗ (3.58) 0.1951∗∗∗ (3.22) −0.0247∗∗∗ (−3.16) −0.0264∗∗∗ (−3.24) −0.0260∗∗∗ (−3.19) 0.0438 (1.51) 0.0310 (1.01) 0.0412 (1.34) MTB S&P 500 Industry effects Yes Yes Yes Year effects Yes Yes Yes Loan purpose effects Yes Yes Yes Loan type effects Yes Yes Yes Observations 7354 6490 6490 66