Lending Relationships and Loan Contract Terms

Lending Relationships and Loan Contract Terms Sreedhar T. Bharath∗ Sandeep Dahiya† Anand Srinivasan§ Anthony Saunders‡ September 11, 2009 Abstract Does repeated borrowing from the same lender affect loan contract terms? We find that such borrowing translates into a 10 to 17 bps lowering of loan spreads. These results hold using multiple approaches (Propensity Score Matching, Instrumental Variables, and Treatment Effects Model) that control for the endogeneity of relationships. We find that relationships are especially valuable when borrower transparency is low and the moral hazard among lending syndicate members is high. We also provide a demarcation line between relationship and transactional lending. We find that spreads charged for relationship loans and non-relationship loans become indistinguishable if the borrower is in the top 30% when ranked by asset size. Similar dissipation of relationship benefits occurs if the borrower has public rated debt or is part of the S&P 500 index. We find that past relationships reduce collateral requirements. Relationships are also associated with shorter debt maturity especially for the lowest quality borrowers. Our results are robust to an estimation methodology which allows loan spread, collateral requirements, and loan maturity to be determined jointly using an instrumental variables approach. We also find relationship borrowers obtain larger loans (scaled by the borrower’s asset size) compared to non-relationship borrowers. Our results imply that, even for firms that have multiple sources of outside financing, borrowing from a prior lender obtains better loan terms. ∗ Stephen M. Ross School of Business, University of Michigan, 701 Tappan Street, Ann Arbor, MI 48109-1234, Tel: (734) 763-0485, E-mail: [email protected] † McDonough School of Business, Georgetown University, G-04 Old North, Washington DC 20057. Tel: (202) 687 3808, E-mail: [email protected] ‡ Stern School of Business, New York University, New York, NY 10012. Tel: (212) 998 0711, E-mail: [email protected] § National University of Singapore, 1 Business Link, BIZ1 Building 04-34, Singapore 117592, Tel: (65) 6516 8434, E-mail: [email protected] 1 Introduction Information asymmetry between lenders and borrowers has played a key role in the development of financial intermediation theory. If an owner/manager cannot credibly reveal the firm’s future prospects, lenders must invest in costly information production and due diligence so as to assess the creditworthiness of potential borrowers and to screen out unacceptably poor quality borrowers (adverse selection). Even if a firm is assessed as an acceptable credit risk, after making the loan the lender must expend resources in monitoring the borrower given the firm’s incentives to invest sub-optimally (borrower moral hazard). However, the information frictions caused by adverse selection and moral hazard can be mitigated if the lending is done by a single private lender such as a bank (Diamond, 1984, Ramakrishnan and Thakor, 1984, Fama, 1985). These risk mitigation benefits are further magnified if the lending bank has had a strong past relationship with a borrower directly producing borrower-specific durable and reusable information (Boot, 2000).1 Although an alternate view, with respect to lending relationships, can be found in Sharpe (1990) and Rajan (1992) who suggest that relationship borrowers may be “locked-in” due to information asymmetries between outside lenders and the borrower. When a loan is shared among multiple lenders, as is the case of syndicated loans, there is an additional element of moral hazard between the lead lender, who is expected to be focal in monitoring the loan (Holmstrom and Tirole, 1997), and other members of the syndicate. Nonlead lenders would likely anticipate less than optimal effort by the lead lender, given that it does not have exposure to the entire loan. We refer to this as “Syndicate Moral Hazard” which results from information asymmetries among lenders and arises due to the lead bank’s incentive to shirk from optimal monitoring. In contrast “Borrower Moral Hazard” results from an informational friction between the lender and borrower due to the borrower’s incentives to divert cash flows for private benefit or to engage in excessive risk-taking. Recent research has begun examining syndicate moral hazard.2 Our paper adds to this literature by examining a large set of bank loans to publicly listed corporations. A key contribution of our paper is to examine the impact of relationships in lowering information asymmetries both between lenders and borrowers as well as between multiple lenders. Specifically, we examine how repeated borrowing from the same lender (which we call relationship borrowing) affects observed loan contract terms. A further contribution of this paper is our estimation of the boundary between relationship and transactional lending. To the best of our knowledge, this is the first paper to estimate the cut-off point beyond which relationship 1 A borrower can benefit from a strong relationship with its lender in a number of ways. These include sharing confidential information such as details of R&D (Bhattacharya and Chiesa, 1995); loan contracts that allow renegotiability (Berlin and Mester, 1992, Boot, Greenbuam, and Thakor, 1993); and the ability to smooth out loan pricing over multiple loans (Berlin and Mester, 1998). 2 Sufi (2007) studies the effect of syndicate moral hazard on syndicate structure. 2 lending becomes indistinguishable from transactional lending in that there are no apparent loan yield/spread benefits to the borrower from past lending relationships. We also examine how past lending relationships affect loan characteristics at different levels of syndicate moral hazard. A number of studies of small, privately held borrowers have provided support for the benefits of relationship lending.3 In this paper, we extend this strand of literature by investigating the role of lending relationships for publicly listed, widely held firms. Our paper differs from previous empirical studies of relationship lending on four dimensions. The first difference is that the firms in our sample have a much wider choice of financing options available, in that all have access to equity markets and a significant fraction to public debt markets. Rapid change in the capital market and increased competition among banks is more likely to impact these firms than private, closely held firms. While Petersen and Rajan (1995) argue that increased competition is likely to erode the benefits of relationship lending, Boot and Thakor (2000) provide a theoretical model which envisages banks engaging in relationship lending as well as “arms-length” transactional lending even as the competition increases. Our sample provides a natural universe to explore the empirical boundary between relationship and transactional lending. The second difference is that our paper focuses on firms borrowing in the syndicated loan market, where a loan is divided among more than one lender. Typically, one or a few lenders (denoted as the Lead Bank(s)) play a delegated monitoring role as described by Diamond (1984). Syndicate members differ in their ability to screen and monitor and the syndicate loan structure results in moral hazard for the lead bank, as it bears all the costs of monitoring the loan, but its share of the loan is less than one hundred percent. Since the monitoring effort of the lead bank is unobservable, the other syndicate members anticipate “shirking” in the level of monitoring effort put in by the lead bank and ex-ante demand higher spreads. Past relationships, which lower the cost of future monitoring, can be seen as a commitment to monitor and can mitigate this syndicate moral hazard problem. Our sample is well suited to test this. The third difference is our use of a large cross-sectional variation in proxies for information opacity regarding the borrower (e.g. size, credit quality, analyst following etc.). We use the term information opacity to capture the idea that higher opacity reflects lower amount of publicly available information. Sufi (2007) provides a useful characterization of opacity “. . . of the degree to which a financial institution must investigate and monitor the borrower.” For example, borrower size (as measured by book value of assets) varies from $93 million at the 25th percentile to $1.6 3 These include Petersen and Rajan (1994), Berger and Udell (1995), Cole (1998) and Degryse and Van Cayseele (2000). The evidence from these studies is mixed. Petersen and Rajan (1994) find that while relationships do increase availability of credit, the interest charged for loans is unaffected. Berger and Udell (1995) report that strong relationships lower both the interest charged as well as collateral requirements. Cole (1998) reports that while the duration of a relationship did not affect the availability of credit, increasing the scope of a relationship as proxied by the purchase of multiple information-sensitive products (such as checking accounts) from a bank, did increase the probability of getting loans from that bank. Degryse and Van Cayselee (2000) find similar results. 3 billion at the 75th percentile. This variation allows us to test how past relationships affect loan contract terms at different levels of borrower information opacity and loan syndicate structure. Fourth, unlike previous relationship papers that have viewed loan contract terms in an independent fashion we allow for joint determination of loan contract terms. Specifically, the joint determination of loan spreads, loan maturity and loan collateral requirements. Equally important, we also evaluate the endogeneity of relationship formation itself using three different approaches - Propensity Score Matching, Instrumental Variables approach, and a Treatment Effects Model approach. We believe that ours is among the first papers in relationship lending that uses all three approaches and documents the advantages and disadvantages of each approach. To preview our results, we find that the benefits of relationship borrowing, measured by a reduction in the loan rate, become insignificant for the top 30% of the firms in our sample ranked by asset size. A similar dissipation of the benefits of relationship lending occurs once a firm has a public debt rating or is part of the S&P 500 index. We find that, on average, repeated borrowing from the same lender is associated with an almost 10-15 basis points (bps) reduction in loan spreads. This reduction is most pronounced for informationally opaque borrowers, consistent with relationships mitigating information asymmetry. We also segregate our loans into groups based on their potential for syndicate moral hazard between lead and non-lead lenders in the syndicate. Our results show that past relationships are significant commitments to monitor, as the presence of such relationships in high moral hazard syndicates is associated with lower spreads. We also estimate the benefits of relationships using the Propensity Score Matching (PSM) methodology. Specifically, this technique uses various observed borrower and loan characteristics to generate an index that measures how likely it is that a loan would be obtained from a relationship lender. Thus, for every relationship loan it is possible to find matching non-relationship loans that had similar propensity to use a relationship lender but did not. The difference in spreads between the relationship loan and matched non-relationship loans allows us to measure the causal impact that relationships have on loan spreads. Our results are remarkably robust, and the difference in spreads ranges from 10 bps to 13 bps across different matching procedures. We also employed two additional methodologies to test for the importance of relationships. The first is an instrumental variables (IV) estimation using distance between the borrower and the lender as an instrument that causes relationship formation. The second, as an alternative to the IV approach, we employ a treatment effects model to control for the endogeneity of relationship formation. Both approaches show significant benefits of relationships on loan spreads, consistent with our PSM findings.. Our results on non-price terms of the loan (collateral and maturity) also find support for the hypothesis that relationships lower information asymmetry between lenders and borrowers. In particular, we find that relationships lower the likelihood that collateral will be pledged (consis- 4 tent with reduction in adverse selection as well as borrower moral hazard). We also find that relationships are associated with shorter maturity of the loan for the lowest quality of borrowers. An interpretation of relationships as credible commitments to monitor is consistent with this finding as shorter maturity provides incentives for more frequent monitoring, which is likely to be less costly for a relationship lender. While the importance of non-price terms in debt contracts has been recognized in previous studies (e.g. Melnik and Plaut, 1986) the empirical evidence has often been limited by a focus on a single contract feature.4 In reality, all the contract terms are likely to be interdependent. We address the econometric issues that arise if the price and non-price terms are determined simultaneously.5 We employ an Instrumental Variables (IV) approach, explicitly recognizing and testing for the endogeneiety of contract terms such as the price, collateral, and maturity of a loan. We find that prior relationships continue to be associated with lower spreads, lower likelihood of collateral, and shorter maturity of the loan (for the lowest-quality borrowers) in the joint specification. Finally, we examine whether access to bank financing is related to borrower-lender relationships. Using the size of a loan facility (scaled by either the borrower’s assets or the borrower’s total long-term debt) as a proxy for access to debt, we find that a firm borrowing from its relationship lender is able to get a loan approximately one to two percent larger than a similar firm borrowing from a non-relationship lender. Our paper thus complements similar evidence for private, closely held borrowers reported by Petersen and Rajan (1994). The remainder of the paper is organized as follows. We discuss the theoretical predictions and testable implications in Section 2. Section 3 describes the data and our sample selection process. Our methodology and major results are presented in Section 4. We conclude in Section 5. 2 Theoretical Predictions and Tests Since our paper is empirical in nature, we first review the theoretical debate surrounding loan spreads, non-price loan characteristics, and relationship lending. In doing so, we highlight how we propose to bring new and additional insights into this debate. As discussed in the introduction, information asymmetries between borrowers and lenders are an important element of financial intermediation theory. Further, in the case of syndicated loans, there is an additional element of moral hazard between the lead lender and other lenders. Below we explain in detail how lending 4 For example Berger et al. (2005) examine the maturity of new loans while Berger and Udell (1995), Jiménez, Salas, and Saurina (2006) focus on the collateral requirements. 5 Dennis, Nandy, and Sharpe (2000) is one of the first papers to address the interrelationship between the price and non-price terms of a loan contract. However, of the four contract features they consider, they analyze only two contract features together at one time. While the features within each set are assumed to be determined jointly, the two sets themselves are assumed to be independent and thus is not a true simultaneous system. 5 relationships can mitigate the adverse selection problem, curb syndicate moral hazard, and lower borrower moral hazard. Boot (2000) argues that relationship lending involves customer-specific information gathered over time through multiple interactions. Further, he argues that that this information must be costly to produce, be proprietary to the lender(s), and be reusable. This definition implies that going back to one’s past lender is likely to reduce adverse selection concerns, since previous transactions would have allowed the lender to generate proprietary inside information about the borrowers. A second effect of relationship lending is to reduce the likelihood of syndicate moral hazard. In particular, not only does the relationship lender have lower information asymmetries with respect to borrower, it also has a lower ongoing cost of monitoring the borrower relative to uninformed outside lenders. As a result it can provide a more credible commitment to monitor the borrower. This commitment to monitor is especially important in syndicated loans where the lead bank does not fully internalize the benefits of its future monitoring effort. Thus, for syndicated loans, the non-lead participant banks can rationally expect a lead bank with a prior relationship to monitor more than a lead bank without a relationship. A third effect of relationships, as a credible commitment to monitor, enables the lender to more effectively control the borrower, once a loan is made. In particular, borrower moral hazard is also mitigated and the borrower is discouraged from engaging in sub-optimal investment and risk-shifting activities. In what follows we develop hypotheses to explore these relationship effects on loan spreads, collateral, maturity and loan amount. 2.1 2.1.1 Effect of relationships on spreads Effect of information asymmetries between lenders and borrowers As argued above, due to lower information asymmetries, the cost of providing future loans (and arguably other information-sensitive services such as securities underwriting) should be lower for a relationship lender. The relationship lender may choose to share, or pass on, these savings to its borrower in a number of ways such as through lower costs of borrowing, more flexible loan contract terms, or a combination of both. If a lender shares some of these benefits with its relationship borrower, then loan costs would be lower for a borrower that uses its relationship lender. Boot and Thakor (1994) also argue that rates charged for loans should decrease as a borrower-lender relationship matures. We call this the relationship benefit effect. An alternate viewpoint regarding lending relationships (Sharpe (1990), Rajan (1992)) posits that the relationship lender may actually distort investment decisions of the borrower and examines the possibility of the borrower being “locked-in.” This lock-in effect would be most relevant 6 only for borrowers’ with few or no alternative sources of financing beyond the relationship bank. Since most firms in our sample are publicly traded, and often have multiple bank relationships, the lock-in effect is likely to be small. However, we do explore the potential for such a lock-in for our sample as well. We employ All-in-Spread Drawn (AISD) as the measure of the interest rate spread charged on a loan. AISD measures the interest rate spread on a loan (over LIBOR) plus any associated fees in originating the loan. Thus, AISD is an all-inclusive measure of loan price. Under the Boot and Thakor (1994) model, relationships should lower spreads. However if the borrower is “informationally captured” (Sharpe, 1990), such a lock-in can result in a relationship lender failing to pass on benefits of lower information production/monitoring costs to its borrower. This would imply that relationship loans need not be significantly cheaper in terms of spreads charged relative to non-relationship loans. Specifically we test the following hypothesis: Hypothesis 1 (H1) Relationship loans, on average, carry a lower All in Spread Drawn (AISD) compared to non-relationship loans. The models of relationships (relationship benefits versus lock-in) also have quite sharp implications regarding the effect of information opacity on spreads. If the relationship benefit effect holds, then the amount of benefit would depend on the relationship lender’s potential to generate proprietary incremental information (and sharing this benefit with its borrower). The benefits may either come from a reduction in adverse selection or lower cost of future monitoring. The greater the information opacity, the greater will be its magnitude. On the other hand, with the lock-in effect, the greater the information opacity, the greater will be the borrower lock-in effect. We use multiple proxies to capture the relative degree of information opacity of a borrower (thus serving as a proxy for the potential to generate proprietary information or potential for lock-in). These include the size of a borrower’s assets, access to public debt markets, inclusion in S&P 500 index, number of analysts that follow the borrower, lagged discretionary accruals, and a microstructure measure of information asymmetry. We test to see whether relationship borrowers faced with higher information asymmetry have a greater reduction in spread (relationship benefit effect) or a lower reduction or even an increase in spread (lock-in effect) compared to non-relationship borrowers. Accordingly we test the following hypothesis: Hypothesis 2 (H2) The higher the information asymmetry faced by a large borrower, the greater the benefits of a relationship on its cost of borrowing. A natural implication of hypothesis 2 is that if borrower transparency is high, the benefits of relationships can become insignificant. For such firms, banks may no longer offer relationship loans. Most theoretical models typically assume that non-relationship or transactional lending is 7 done in the capital market, while relationship loans are assumed to be provided by banks (Rajan, 1992). However, in reality banks engage in both relationship lending as well as transactional lending. Boot and Thakor (2000) describe a model in which the banks and capital markets co-exist, and banks may engage in relationship or transactional loans. A central result of their model is that there exists a critical level of borrower quality below which a bank would choose to provide relationship loans. For higher-quality borrowers, the bank would offer transactional loans. If relationship loans are marked by lower spreads as argued in our hypothesis 1 above, and the effect is greater for firms that are more informationally opaque as implied by hypothesis 2, Boot and Thakor’s (2000) model implies a third possible way to test for the impact of informational asymmetries among lenders and borrowers. In particular, we can estimate the critical point, in terms of borrower quality (i.e., observable measures of risk such as credit ratings and firm size), where there is a switch from relationship lending to transactional lending. The various measures of information opacity outlined earlier allow us to estimate the delineation point across different borrower characteristics (such as size) where the benefits of relationship borrowing (if any) become insignificant. We test the following hypothesis, that is conditional on finding a net relationship benefit effect. Hypothesis 3 (H3) There exists a borrower quality above which the benefits of relationships on loan rates are insignificant. 2.1.2 Additional effect of syndicate moral hazard Given that a large fraction of our loans are syndicated among multiple lenders, our data allows us to study the interaction of past lending relationships with moral hazard among multiple lenders. Holmstrom and Tirole (1997) show that in a setting that involves multiple lenders and where one lender (of a few) is expected to do the monitoring, the monitoring lender faces significant moral hazard. The intuitive idea being that the monitoring bank bears all the costs but shares only part of the benefits from engaging in such due-diligence, and would rationally shirk from putting in the optimum effort. Ex-ante, other lenders take these incentives into account and demand a higher rate to compensate them for the anticipated shirking by the monitoring bank.6 Sufi (2007) tests the above model and finds that loan syndicates are structured to minimize this moral hazard problem. In particular, he finds that lending syndicates tend to be more concentrated and the lead bank retains a higher share of the loan for borrowers requiring high levels of monitoring. Thus, observed syndicate structure is a reasonable proxy for the syndicate moral hazard faced by the lead bank. We use the two measures of moral hazard used by Sufi 6 A somewhat similar issue related to multiple lenders is discussed by Diamond (2004). He shows that for an appropriately structured debt contact, a large syndicate commits each lender in the syndicate to enforce the contract even when such enforcement is costly. 8 (2007), namely share of loan retained by the lead bank and Herfindahal index of loan shares. In addition, we also use the size of a syndicate to proxy for moral hazard. If a past lending relationship can be viewed as a commitment to monitor diligently on future loans (and this commitment arises because the relationship bank faces a lower ongoing cost of monitoring than a comparable non-relationship bank), its presence would alleviate the moral hazard concerns of the syndicate participants, resulting in lower rates for such loans. Specifically, we test the following hypothesis: Hypothesis 4 (H4) Relationship loans would have a lower spread compared to non-relationship loans in syndicates with high moral hazard. 2.2 Effect of relationships on non-price terms Next, we focus on how non-price terms are affected by lending relationships. Collateral and duration (maturity) of a loan are considered key loan contract features. We discuss collateral first. The use of collateral in debt contracts has been justified on two grounds: adverse selection and borrower moral hazard. While both arguments rely on information asymmetries, they provide opposite predictions of what type of borrowers would post collateral. The adverse selection models (Bester, 1985, and Besanko and Thakor, 1987) argue that willingness to provide collateral serves as a credible signal of borrower quality. These models predict that better-quality (i.e. low credit risk) borrowers would post collateral and obtain lower spreads for the loans. These models suggest that collateral and interest rate are substitute mechanisms. Borrower moral hazard models (Holmstrom and Tirole, 1997, Stultz and Johnson, 1985, and Boot et al., 1991) stress the ex-ante incentives for asset-substitution when firms take on risky debt. Holmstrom and Tirole sum it up concisely “. . . Firms with low net worth have to turn to financial intermediaries, who can reduce the demand for collateral by monitoring more intensively. Thus, monitoring is a partial substitute for collateral . . . (p. 665).” These incentives are strongest for the informationally opaque (presumably perceived as high credit risk) borrowers. In equilibrium, such borrowers can credibly commit to lower asset-substitution by providing collateral. Thus, these models predict that the lowest-quality borrowers are more likely to be required to provide collateral. However, the focus of our paper is not on the determinants of collateral per se, but rather we focus on the impact of relationships on the requirement of collateral. To the extent that relationships reduce adverse selection problems due to lower information asymmetry, the relationship lender would not require collateral from their borrowers since they are not subject to this problem. On the other hand, if the borrower moral hazard view of collateral is true (again, viewing relationships as a commitment to monitor), the relationship bank would not require as much collateral as 9 an equivalent non-relationship bank.7 Thus, the effect of relationships on collateral is to imply a lower collateral requirement, both due to reduction of adverse selection and due to a higher commitment to monitor. This yields our next hypothesis: Hypothesis 5 (H5) The probability of using collateral as a loan contract non-price term decreases if the loan is provided by a relationship lender. We also examine the maturity structure of loan contracts. While Flannery (1986) suggests that debt maturity would increase as borrower quality improves, Diamond (1991) predicts that corporate debt maturity would exhibit a non-monotonic relationship with borrower quality. One of the central results of Flannery (1986) is that if transaction costs of debt issuance are high enough, borrowers with good future prospects can credibly convey their unobservable quality via choice of their debt maturity. Thus, Flannery’s model predicts a linear relationship between borrower’s unobservable credit quality and debt maturity - better are the (unobservable) future prospects of a firm, the shorter is its the debt maturity structure. Diamond uses a trade-off between information asymmetries and the liquidity risk of refinancing. On one hand, frequent refinancing necessitated by shorter maturity makes borrowing costs more sensitive to information which is desirable for high-quality (high credit rating) firms. On the other hand, shorter maturity exposes a borrower to liquidity risk, that is, a temporary shock which can either cause the price of loans to go higher or financing can become difficult to obtain. Diamond summarizes the key insight of his model as “. . . For a borrower with a sufficiently good credit rating, this liquidity risk is outweighed by the effect of expecting future news to be favorable. For borrowers with lower ratings, the liquidity risk outweighs the information effect . . . However, very low rated borrowers may have no choice but to choose short-term debt, despite the incentives for inefficient liquidation that it gives to lenders. The two types of borrowers that choose primarily short-term debt imply that the chosen debt maturity is not a monotonic function of the borrower’s credit rating . . . empirical studies of maturity will measure a mixture of two effects. This could make inferences complicated . . . ” Based on the potential costs of a liquidity shock, Diamond’s model predicts that the best quality firms (with high credit ratings) would demand short-term maturity debt while intermediate quality firms would opt for long-term debt. The lowest quality borrowers require intense monitoring and would thus be supplied with only short-term loans. How do borrowing relationships interact with the effect described above - namely information asymmetry, liquidity risk, as well as the demand-supply effects for higher and lower quality firms? In the model described by Flannery, relationships would lower information asymmetries. As the need to signal quality through debt maturity becomes diminished due to better information sharing, strong relationships should result in longer maturity loans. In Diamond’s model such 7 Jiménez et al. (2006) provide an extensive overview of this literature and provide empirical evidence which that shows that lower credit quality firms are significantly more likely to face collateral requirements. 10 a reduction in adverse selection problem is only relevant for the high- and intermediate quality borrowers, since debt maturity for these borrowers is largely demand driven. For high and medium quality borrowers, the reduction in information asymmetries due to past relationships leads to two opposing effects: On one hand, better information sharing removes the need to borrow short term to obtain favorable interest rates by the high quality firms, thus predicting longer maturity for such borrowers. On the other hand such reduction in information asymmetries also reduces the possibility of inefficient liquidation which makes the shorter-term debt more appealing for both high and medium quality firms. For the lowest quality firms, however, debt maturity is largely determined by supply factors. For these borrowers Diamond’s model predicts shorter debt maturity to enable more frequent monitoring by the lender. Past relationships reduce the cost of such frequent monitoring. Thus for low quality borrowers, relationship lenders would offer shorter maturity since they face a lower monitoring cost. To summarize, the effects of relationships on debt maturity are likely to differ across different classes of borrower quality. This is captured in our next hypothesis: Hypothesis 6 (H6) For low quality firms, relationship banks will commit to monitoring more by providing shorter maturity loans. For medium- and high-quality firms, relationships may either increase or decrease maturity. We also examine issues relating to the simultaneity of various loan contract terms in evaluating H5 and H6. Arguably, the interest rate (plus fees for AISD), collateral, and maturity on a loan are determined simultaneously. Most studies on debt contract terms have not considered simultaneity issues directly. We re-estimate the relationship between the price and non-price terms of loans using an Instrumental Variables (IV) framework. This approach requires identification of relevant and valid instruments for collateral, maturity, and spreads. We employ a number of possible instruments to estimate our system and discuss them in section 4.9. A number of studies have documented the positive association between the length of a banking relationship and availability of credit to a relationship borrower (Petersen and Rajan, 1994 and Degryse and Van Cayseele, 2000). Hoshi et al. (1990) report that for a sample of financially distressed Japanese firms, those with strong bank relationships were less likely to be credit constrained compared to the borrowers lacking strong banking relationships.8 These empirical findings provide support to the argument that a strong banking relationship is likely to increase access to financing for a borrower. As such, relationships reduce lender-borrower information asymmetries and likelihood of syndicate moral hazard. We employ the size of loan facility (which we scale by either the borrower’s assets or its existing long-term debt) as a proxy for access to financing. We formalize this benefit in our next hypothesis: 8 In a recent study, Faulkender and Petersen (2006) report that even large publicly listed firms face credit constraints - less transparent firms (defined as firms lacking public debt rating) have leverage levels almost 30% lower compared to more transparent borrowers. 11 Hypothesis 7 (H7) A borrower with a prior relationship with its lender would be able to obtain a larger loan amount compared to a similar borrower that does not have a prior relationship with its lender. 3 Data and Sample Selection Data on individual loan facilities is obtained from the Dealscan database maintained by the Loan Pricing Corporation (henceforth, LPC).9 LPC has been collecting information on loans to large U.S. corporations primarily through self-reporting by lenders, SEC filings, and its staff reporters. Strahan (1999) provides a good description of the LPC Dealscan database. While the LPC database provides comprehensive information on loan contract terms (LIBOR spread, maturity, collateral, etc.), it does not provide much information on borrowers. We manually match the borrowers in the LPC database with the merged CRSP and Compustat database following the procedure outlined in Bharath et al. (2007). We then use Compustat to extract data on accounting variables for the given company. We also extract the primary SIC code for the borrowers from Compustat and exclude all financial services firms (SIC codes between 6000 and 6999). To ensure that we only use accounting information that is publicly available at the time of a loan we employed the following procedure: For those loans made in calendar year t, if the loan activation date is 6 months or later than the fiscal year ending month in calendar year t, we use the data of that fiscal year. If the loan activation date is less than 6 months after the fiscal year ending month, we use the data from the fiscal year ending in calendar year t-1. Our sample period is from 1986 to 2003. Over this period there was extensive mergers and acquisitions activity in the U.S. banking sector. To construct a chronology of banking mergers/acquisitions, we used the Federal Reserve’s National Information Center database and complemented it by hand matching the data from the SDC mergers and acquisition database, LexisNexis, and the Hoover’s corporate histories database. This allows us to trace lending relationships through time even if the original relationship lender disappears due to a merger or an acquisition. To examine the impact that prior lending relationships have on the price and other terms of loan, we need to segregate our loans into those that are provided by a relationship lender and those that are provided by a non-relationship lender. Following Dahiya, Saunders, and Srinivasan (2003) we construct the relationship measures for a particular loan i by searching all the previous loans (over the previous 5-year window) of that borrower as recorded in the LPC database. We note the identity of all the lead banks on these prior loans and if at least one of the lead banks for loan i had been a lead lender in the past we classify loan i as a relationship loan. Since the identification of the “lead” bank (or banks) for a particular loan facility is the basic building block 9 LPC Dealscan database collects the data on loans made to large (mostly publicly traded) U.S. firms. It has been widely employed to study the private debt market. See, for example, Drucker and Puri (2005). 12 of classifying a loan as relationship or non-relationship, we present a detailed discussion on this identification below. LPC’s Web-based product called LoanConnector provides a field labeled called ‘Lead Arranger Credit’ which can take values of ‘Yes’ or ‘No’ for every bank. The banks that are classified as a lead arranger credit hold a large fraction of the loan, on average about 58.88% over the entire sample period. In our examination of the data we find that for most loans it is a single bank that is accorded the role of lead arranger credit by LPC. Thus, we classify such banks as lead banks. This mitigates the possibility that we may misclassify banks that are not relationship banks but make a large loan commitment. Sufi (2007) uses the field lead arranger credit as the sole criterion for classifying a bank as a lead bank. However, we need to account for the possibility that other roles may also represent banks that are also truly in a lead position. Therefore, we examine loan shares retained by banks in all the 38 distinct roles in our sample. We identified the following roles where the bank retained a significant share of the loan (greater than 25%): Agent, Administrative Agent, Arranger, Lead Bank.10 We designate banks that were retained in any of the above four roles as lead banks. Finally, sole lender transactions by construction have a clearly identified lead bank and we designate it as such. Thus our classification of what constitutes a lead bank is somewhat broader relative to Sufi (2007). For close to 90% of loans in the sample, our methodology results in a single bank being classified as the lead bank. We believe this approach mitigates the problem of misclassification of relationship lenders.11 Next, for every facility, we construct three alternative measures of relationship strength by looking back and searching the past borrowing record of the borrower. We search the previous 5 years by starting from the activation date of that loan facility.12 The relationship variable is denoted by REL(M), where M is one of the three alternatives measures. In the following paragraphs, we describe the process of constructing these relationship measures using a set of loans obtained by Owens Corning, one of the borrowers in our sample. In June 1997, Owens Corning borrowed $2 billion from a syndicate of banks. The lead bank for the June 1997 issue was Credit Suisse. It was the only bank that was given lead-arranger credit for this deal. In addition, it was the only bank retained in one of the four roles that was 10 The roles have evolved over time. In the early part of the sample “Agent” was frequently used for the lead bank, in the later years “Arranger” and “Administrative Agent” have become more common. 11 We see some signs that syndication is becoming a specialist job, separable from being a relationship bank. We did find a new role titled “Syndications Agent” that appears frequently toward the end of the sample period but does not appear at all in the initial years. We exclude this role from the computation of our relationship measure. 12 To be classified as either a relationship or a non-relationship loan, we require that there be at least one loan in the 5-year window prior to the loan being classified. If this condition is not met, we exclude the given loan from the sample of classified loans. This approach also implies that the first loan of any firm in the LPC database is not included in our analysis since by construction there can be no prior loans to allow the classification of the first loan as either relationship or non-relationship. 13 considered to be a lead position. To estimate REL(M) for this $2 billion loan, we look-back on the previous 5 years of borrowing history of Owens Corning up until the date of this loan to see if the lead bank of the current facility had provided loans in the past 5 years. LPC reported three loans taken by Owens Corning over this 5-year look-back period. In September 1993, it borrowed $475 million from a syndicate where Credit Suisse was the only bank retained that was given the lead-arranger credit and retained in the role of an agent. It borrowed another $110 million from a syndicate led by Bank of New York in May 1994, where Bank of New York was given the lead-arranger credit and retained in the role of an agent. Then, in December 1995, it borrowed $99.6 million where Credit Suisse was given lead-arranger credit and retained in the role of an agent. Thus, looking back from the point of the June 1997 loan, Owens Corning contracted three loans totaling $684.6 million (475+110+99.6) in the 5-year look-back period. Next, we describe the construction of various relationship measures for the June 1997 loan. The first relationship strength variable is a binary measure of relationship and is designed to pick up the existence of prior lending by the same lender in the past. For a particular bank m it is denoted by - REL(Dummy)m . In our example, REL(Dummy) would equal 1 for Credit Suisse. In case there were multiple lead banks retained, we pick the highest value of REL(Dummy) for all the lead banks and assign it to the loan. In our example, the June 1997 loan has REL(Dummy) =1. The other two measures of relationship strength are continuous. The first continuous measure of relationship strength is REL(Amount). For bank m lending to borrower i, it is calculated as REL(Amount)m = $ Amount of loans by bank m to borrower i in the last 5 years Total $ amount of loans by borrower i in the last 5 years (1) Again, this measure is calculated for each of the lead banks and the highest value across all lead banks is used in our analysis. Thus, in case of June 1997 loan to Owens Corning REL(Amount)Credit ). is 0.84 ( 475+99.6 684.6 The second continuous measure of relationship strength is REL(Number). For bank m lending to borrower i, it is calculated as REL(N umber)m = Number of loans by bank m to borrower i in last 5 years Total Number of loans by borrower i in in last 5 years (2) In our example, REL(Number) would be 0.67 (calculated by dividing 2 by 3) for Credit Suisse. For loans with multiple lead banks, the highest REL(Number) is used.13 13 To illustrate the process of estimating REL for a particular facility with multiple lead lenders, consider a loan facility that we need to assign a relationship measure to. Assume that this loan facility has two lead banks: bank A and bank B. To estimate whether this facility is a relationship loan we first check the previous 5 years to see if either bank A or bank B has been a lender in the past. If this condition is true REL(Dummy) for this facility is assigned the value one and zero otherwise. Estimation of REL(Number) (and REL(Amount)) requires that we 14 Suisse Table 1 provides the descriptive statistics for our data and segregates relationship and nonrelationship loans (i.e., loans from a bank that did not have a past relationship with the borrower). Panel A provides the calendar-time distribution of the loan sample. The low number of observations in the early years is largely due to improvement in coverage in the LPC database over time. Panel B illustrates the one-digit SIC classification of the borrowers. There is a strong concentration of loans in the manufacturing sector (SIC codes between 2000-3999). Panel C lists the primary purpose of the loan facility contracted, with loans for corporate purposes and debt repayment the most frequently reported purposes. Following Drucker and Puri (2005), we use the LPC reported “All-in-Spread-Drawn” (hereafter AISD) as the measure of interest rate for a loan. AISD is the coupon spread over LIBOR on the drawn amount plus the annual fee. Since the interest charged for a loan is affected by various loan-specific characteristics (maturity, loan size, etc.) and borrower-specific characteristics (borrower size, profitability, leverage, etc.), we obtain these variables from LPC and COMPUSTAT respectively. Table 2 reports the various sample summary statistics. The data is winsorized at the one percent and 99 percent level to address the problem of extreme outliers. The median AISD is 212.5 bps. The median loan facility is $50 million. The median book value of assets for our sample of borrowers is $361 million. The high fraction of syndicated loans (79%) reflects the historical focus of LPC on collecting data on large syndicated loans. The average maturity for loan facilities is 43 months (median 36 months). 4 4.1 Methodology and Results Univariate Tests of H1 (Loan Spreads) To examine if repeated borrowing from the same lender(s) affects loan contract terms, we first examine key loan features to see if these are significantly different for relationship vs. nonrelationship loans to the borrowers. In Panel A of Table 3, we segregate the entire sample based on the existence of prior relationships to test if the loan contract terms reflect prior lending relationships. In the first column, we report key loan terms for loans taken from non-relationship lenders. The second column provides the same information for relationship loans. These loan terms include loan price variables: all-in-spread-drawn (AISD), all-in-spread-undrawn (AISU), upfront fee, and annual fee.14 estimate it for both banks. Suppose bank A has been the lead in 1 out of 3 loan facilities this borrower obtained in the previous 5 years, it would imply that REL(Number) is 1/3 = 0.333. Now we do the same calculations for bank B - it turns out that bank B has been the lead in 2 out of the same 3 loans. In that case REL(Number) is 2/3= 0.667. We assign the higher value (0.667) as the REL(number) for that facility. 14 AISU is the spread paid on the undrawn loan amount. 15 Other loan-specific features reported include loan facility size, maturity, and collateral (denoted as percentage of facilities secured). The last column reports the differences in mean (median) loan characteristics between relationship and non-relationship loans. The results of univariate tests of differences in means and medians provide strong evidence that relationship loans enjoy significantly better loan price terms as well as non-price terms such as loan size and collateral requirement. Comparing AISD (the most comprehensive measure of borrowing cost) for a company borrowing from a relationship lender, we find that on average AISD is 52 basis points lower compared to a borrower that does not have a prior relationship with the lender. This difference is significant at the one percent level. AISU, up-front fees, and annual fees are all lower for relationship borrowers, and the difference is significant at the one percent level. Results for non-price loan terms show similar effects. Typically, loans to relationship borrowers are less likely to be secured and are larger on average. Again these differences are significant at the one percent level and economically large. Thus, the price effects documented in the univariate tests are more consistent with the relationship benefit effect than “lock-in” effects. While the univariate tests provide preliminary evidence that borrowers derive significant benefits from having strong relationships with their lenders, these results do not take into account potentially significant differences in borrower characteristics between the relationship borrower and non-relationship borrower groups. It is likely that the relationship borrowers have fundamentally different characteristics. For example, banks may prefer to maintain relationships with borrowers with a track record of strong financial performance. To determine if borrowers using relationship lenders obtain better loan terms, we must first test whether the characteristics of the two groups (relationship and non-relationship borrowers) are different, and whether these differences explain the difference observed in loan features across these two groups. We compare the borrower characteristics in the two groups. The results are reported in Panel B of Table 3. The average size (as measured by book value of assets) of a relationship borrower ($ 4,075 million) is almost twice the average size of a non-relationship borrower ($ 2,126 million). This difference in size is significant at the one percent level. Firms borrowing from relationship banks also differ from those borrowing from non-relationship banks across measures of leverage and profitability. For example, firms borrowing from relationship lenders have a higher EBITDA to sales ratio (16% versus 14%), higher long-term debt to assets ratio (27% versus 25%), and lower current ratios (1.90 versus 1.96). The higher leverage and lower current ratios of relationship borrowers suggest that these borrowers have better access to bank loans. Relationship loans are also more likely to have a credit rating and more likely to have an investment grade rating. These differences between the two borrower groups are statistically significant at the one percent level. Tests for difference in medians provide qualitatively similar results. While the results of univariate tests suggest that firms benefit significantly by borrowing from 16 relationship banks, these results also show that some of the key borrower characteristics that influence the cost of loans are systematically different across the relationship and non-relationship borrowers. Consequently, to better distinguish between relationship and performance effects on borrowing costs we employ multivariate tests. 4.2 Multivariate Tests of H1 (Loan Spreads) Since the cost of borrowing is likely to depend on various loan-specific features such as loan size, maturity etc., as well as on a borrower’s historical performance, we use a regression model of the following form. X AISD = β0 + β1 (REL(M )) + βi (Loan Characteristicsi ) X X βk (Controlk ). + βj (Borrower Characteristicsj ) + (3) The variables are defined below: • AISD: The dependent variable is ‘All In Spread-Drawn” (AISD), which equals the coupon spread over LIBOR on the drawn amount plus the annual fee. • REL(M): This is the measure of relationship strength constructed by looking back and searching the past borrowing record of the borrower. As discussed earlier, we construct 3 different specifications for this variable. • Loan Charactersticsi : Various characteristics of loan facility as described below : – LOG(MATURITY): The natural log of maturity of loan facility in months. – LOG(LOAN SIZE): The natural log of loan facility amount adjusted for inflation in year 2000 dollars. – COLLATERAL: A dummy variable that equals 1 if the loan facility was secured and 0 otherwise. • Borrower Characteristicsj : Various characteristics of the borrower as described below: – LOG(ASSETS): The natural log of the book value of the assets of the borrower adjusted for inflation in year 2000 dollars. This controls for cross-sectional variation in borrower size in our sample. – LEVERAGE: Ratio of book value of total debt to book value of assets. EBIT DA ). – LOG(1+COVERAGE): Calculated as natural log of ratio (1+ Interest Expenses 17 – PROFITABILITY: Ratio of EBITDA to Sales. – TANGIBILITY: Ratio of Property, Plant, and Equipment (PPE) to total assets. – CURRENT RATIO: Ratio of current assets to current liabilities. – MARKET TO BOOK: Calculated as ratio of (book value of assets-book value of equity + market value of equity) to book value of assets. • Controlk : These are other control variables and include dummy variables for the year of the loan facility, loan purpose, loan type, S&P senior unsecured debt rating with not rated firms considered as a separate group, and the industry of the borrower (one-digit SIC code). Results of this regression (equation 3) using REL(M) as the relationship measure are reported in Table 4. Regardless of which measure is used, the coefficient on the relationship variable is negative and significant at the one percent level. Standard errors used to assess significance are corrected for heteroscedasticity and firm level clustering (Rogers, 1993).15 Holding all else constant, the cost of borrowing from a relationship lender is lower by almost 10 basis points (bps) compared to borrowing from a non-relationship lender. Given our univariate tests that show an approximately 50 basis point difference between relationship and non-relationship borrowers, the multivariate result implies that relationship variable alone accounts for roughly 20% of that difference. As foreshadowed in our univariate results, the cross-sectional differences in borrower characteristics (relationship versus non-relationship) explains much of the variation observed in spreads charged by lenders. For example, holding all else constant a borrower at the 75th percentile of size (as measured by book value of assets) would on average pay approximately 70 bps lower on a similar loan compared to a borrower at the 25th percentile of size. As expected, lower leverage, higher profitability, higher current ratio are associated with significantly lower spreads charged on loans. Interestingly, the negative (and significant) coefficient for maturity and positive (significant) coefficient for collateral are inconsistent with the notion that these non-price terms can be used as trade-off features for price terms. These results, however, conform with those reported by Berger and Udell (1990) who also find that borrowers that are required to post collateral are also more likely to be paying higher spreads. In the first three columns we use year dummies to control for calendar time effect, and in column 4, instead of year dummies we include the prevailing market default spread at the time of the loan. The default spread is measured as the difference between the yield on Moody’s seasoned corporate bonds with Baa rating and 10-year U.S. government bonds. In this specification, the coefficient for REL(Dummy) is -11.15 and is significant at the one percent level. The coefficient 15 Some loan facilities make-up a single deal. Our analysis is done at the facility level. We also verified that clustering at the deal level produces similar results. 18 for default spread prevailing at the time of the loan is 34.47 and is also significant (at the 1% level). 4.2.1 Other Robustness Tests We check for robustness of our results by re-estimating our main specification (equation 3) in different ways. First, we drop all loan facilities classified as 364 day facilities in calculating REL(Dummy).16 These facilities arguably exist only because of bank capital regulation, and the set of borrowers that use them may be systematically different than other borrowers. The coefficient for the revised REL(Dummy) is now -10.25 (significant at the one percent level). This estimate is virtually identical to the original estimate of -10.55. In another exercise, we drop all loan facilities which appear to be amendments of existing facilities, likely to be initiated by a borrower with improving credit prospects.17 The coefficient for REL(Dummy) is now -12.80 which is significant at the 1 percent level.18 Overall, the multivariate results show that even after controlling for various loan, market, and borrower-specific characteristics, borrowing from relationship lenders is associated with significantly lower spreads, consistent with relationships providing benefits to borrowers.19 16 All our results are robust to using REL(Number) or REL(Amount) in these tests as well as all subsequent tests in the paper. To conserve space we do not report these results in detail but these are available on request. 17 We searched the amendment comment field of every loan facility in the DealScan database to check if the facility was an amended facility. 18 These robustness results are available on request. 19 A possible alternative explanation of our findings is that observed reduction in spreads simply reflects the lower transaction costs of lending to the same firm. If transaction costs are the sole determinant of relationship cost savings, then these savings should depend on the time elapsed between the current loan and the previous loan. Arguably, transaction cost savings are highest for loans that are made frequently and when the time between successive loans is short. Thus, the benefits of repeated borrowing should dissipate as the time from the first loan increases or if the loans are infrequent. For each loan in our sample we create two measures to capture this idea of time between loans and frequency of loans. “LOAN INDEX” equals the number of loans obtained by the borrower prior to the current loan. “TIME TO LAST LOAN” is the time measured in months between the current loan and the most recent loan. If the benefit of repeated borrowing from the same lender is due to reduced transaction costs, we should expect the benefits to decline as both Loan Index and Time to Last Loan increase. Specifically if we interact the REL and one of these measures of time between loans, we should expect a positive and significant coefficient for the interaction term. On the other hand if these benefits represent an enduring information advantage due to close relationships, the interaction term should be insignificant. To conserve space we do not report the results of these specifications. We find that interaction term is insignificant for both Loan Index and Time to Last Loan. This suggests that the benefits of lower spreads for repeated borrowing from the same lender are unlikely to be driven solely by lower transaction costs. 19 4.3 4.3.1 Endogeneity of Relationship Formation Propensity Score Matching Tests A basic drawback of our relationship measure is that the decision to form a relationship or to break a relationship may be endogenous. The decision to form and stay in a relationship is to a certain extent determined by the borrower, which in turn is likely to be related to observed firm characteristics such as borrower size. Ideally, one would like to run an experiment with pairs of matched firms which that are identical in all respects except relationships. One firm in each pair would borrow from a relationship lender while other borrows from a non-relationship lender. The observed difference in loan rates across all pairs would then be a robust estimate of the effect of past relationships on loan rates. While such an experiment is not feasible, econometric techniques can provide good matched samples based on observable characteristics. We employ the Propensity Score Matching (PSM) technique described by Heckman et al. (1997, 1998). This methodology has been used by Drucker and Puri (2005) among others in recent studies. Essentially, this technique estimates the predicted probability of group membership (e.g., probability of being in the treatment group versus the control group in a clinical drug trial) based on observed characteristics using a probit model. In our case, for each loan we use a number of loan and borrower characteristics to generate the probability of that loan being obtained from a relationship lender. Specifically we estimate a probit model of the following form: X REL = β0 + βi (Loan Characteristicsi ) + X X βk (Controlk ). + βj (Borrower Characteristicsj ) + (4) The dependent variable is REL, a dummy variable that equals one if there is a past relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise. The loan characteristics include log of loan size, and dummy variables for the type and purpose of the loan. Borrower characteristics include log of assets, profitability, tangibility, leverage, interest coverage, current ratio, market to book ratio, and borrower rating. Other controls include onedigit SIC code of the borrower, term spread, and default spreads prevailing at the time of the loan origination. For each loan, we estimate predicted probability (i.e., propensity score) of it being a relationship loan. We then match each relationship loan with a set of non-relationship loans that have propensity scores similar to that of the relationship loan. Heckman, Ichimura, and Todd (1997, 1998) describe the matched estimators we use. The NEAREST NEIGHBOR estimator chooses for each relationship loan, the n loans with closest propensity scores and uses the arithmetic average of the AISD of n non-relationship loans (We use n = 10 and n = 50). The GAUSSIAN and EPANECHNIKOV estimators use a weighted average of non-relationship 20 loans, with more weight given to non-relationship loans with propensity scores that are closer to the relationship loan propensity scores. The GAUSSIAN estimator uses all non-relationship loans, while for the EPANECHNIKOV estimator, we specify a propensity score bandwidth (h) that limits the sample of non-relationship loans to be used for comparison. We specify that h = 0.01. We report our results in Table 5, Panel A. In column (1), we compute mean AISD difference between relationship loans and non-relationship loans by using the propensity score estimators to match them. Across all four matching specifications (NEAREST NEIGHBOR (n=10), NEAREST NEIGHBOR (n=50), GAUSSIAN and EPANECHNIKOV) we find that spreads on relationship loans are 11 to 12 bps lower. These differences are significant at the one percent level (t-ratios are estimated using standard errors obtained by bootstrapping with 50 replications). While column 1 provides estimates of relationship effects that are robust to endogeneity issues, we also need to address an additional concern. The group of loans we classify as non-relationship loans includes two distinct subgroups. The first subgroup includes borrowers that had a prior lending relationship, but for the loan being classified, chose (or were forced) to obtain it from a non-relationship lender. Such a loan denotes a break-up of an existing relationship. A second subgroup consists of loans where the borrower never formed a relationship in the past. Arguably these two types of non-relationship loans may carry significantly different loan rates. For example, the decision to switch or break up an existing relationship presents the new lender with an adverse selection problem (Detragiache, Garella, and Guiso, 2000) and may result in higher loan spreads. To examine this issue in more detail, we divide our sample of non-relationship loans into “break ups” and “non-break ups.” For every non-relationship loan, we look back on the most recent loan. If the most recent loan was a relationship loan, we classify the loan currently being examined as a break up. If the prior loan was also a non-relationship loan, we classify the loan currently being examined as non-break up.20 In column (2), we compute mean AISD difference between relationship loans and only those non-relationship loans where the borrower had a relationship but chooses to obtain the loan from a non-relationship lender, thus breaking its existing relationship. In column (3), we compute mean AISD difference between relationship loans and only those non-relationship loans where the borrowers did not have a past relationship. The benefit of maintaining a relationship versus breaking it is almost 13 bps and significant at the one percent level across all 4 specifications (column 2). The difference in spreads from borrowers who choose to stay in relationships compared to those borrowers who do not form relationships is lower (about 8 bps) and also significant (column 3). 20 This methodology implies that there would be some non-relationship loans that cannot be classified into these two sub-groups because the most recent loan is the first loan for the borrower in the LPC database and is thus unclassified. We exclude such loans from our analysis. 21 Check for Reliability of the Propensity Score Matching Results Propensity Score Matching estimators are not consistent estimators for treatment effects if the assignment to treatment is endogenous, i.e., if unobserved variables that affect the assignment process are also related to the outcomes. Also, the matching method is based on the conditional independence or unconfoundedness assumption, which states that the researcher should observe all variables simultaneously influencing the participation decision (propensity to form relationships, the treatment) and outcome variables (spreads). This is a strong identifying assumption and should be justified. If there are unobserved variables that simultaneously affect assignment into the treatment and the outcome variable, a hidden bias might arise to which matching estimators are not robust (Rosenbaum, 2002). In order to estimate the extent to which such “selection on unobservables” may bias our qualitative and quantitative inferences about the effects of relationships on loan spreads, we conducted a sensitivity analysis as outlined in Rosenbaum (2002). The details of this analysis are described in the Appendix. Overall, results of this sensitivity analysis suggest that selection on unobservables is unlikely to weaken our results. 4.3.2 Instrumental Variables Estimation A potential source of endogeneity is that there may be a common unobserved factor that drives both the formation of a relationship as well as the loan spread. Relationship is the main variable of interest in our empirical model where we seek to explain the cross-sectional variation in observed loan spreads (AISD). To the extent that there are other borrower specific characteristics that we do not control for that may explain both AISD as well as the existence of relationship, our coefficient estimates are potentially biased. Unobserved credit quality is a possible factor, in the sense that a bank might tend to form relationships with firms of high credit quality (unobservable to the econometrician). Thus, lower loan spreads on relationship loans might not be because of incremental benefits of relationship lending as we have argued so far. It might simply be the result of the relationship variable proxying for unobservable higher credit quality of the borrower. One potential solution is to use an instrument that is correlated with the relationship formation but does not affect the loan spreads directly except through relationships. We use the geographic distance between the borrower and its lead lender as an instrument for relationships. A number of papers have argued that information gathering and processing (which are at the heart of relationship lending theories) is more easily done when the physical distance between a lender and a borrower is shorter. For example, Berger, Miller, Petersen, Rajan, and Stein (2005) describe this succinctly: “. . . being close to one’s customers is likely to facilitate a loan officer’s collection of soft information . . . Being nearby might also help the loan officer to better understand the nuances of local business environment . . . ” (Pg. 244). Coval and Moskovitz (2001) find that mutual fund managers’ investment in firms that are physically close to the fund manager generate 22 significantly higher returns and argue that this may arise due to better information: “Investors located near potential investments may have significant information advantages relative to the rest of the market . . . ” (Pg. 839). Dass and Massa (2006) also use the same argument of better information gathering/processing for physically closer borrower and cast their lender-borrower relationship in terms of geographical distance. These papers suggest that distance is likely to be correlated with propensity to form relationships. Other than its effect through the lending relationship, location should not have any impact on spreads.21 For each borrower, information on that firm’s headquarters, i.e., city, state, and zip code is taken from Compustat. For each loan we locate the city in which the borrower has its headquarters and find the latitude and longitude of the city. We do a similar exercise for headquarters of the lead bank. With the longitude and latitude data for both the lender and the borrower, one can calculate the spherical distance between the two.22 As discussed in Petersen and Rajan (2002), we use log(1+distance) to address the skewness in the distance variable. In the first stage regression, we examine the determinants of a loan being a relationship or a non-relationship loan. The results of this first stage probit model are provided in Panel B of Table 5. We argue that the propensity of forming a relationship would decrease as the physical distance between the borrower and the lender increases. Consistent with this, we find that coefficient for log(1+distance) is negative (0.058) and significant at the one percent level. These results suggest that log(1+distance) appears to be sufficiently correlated with relationships formation to be a viable instrument. Since the dependent variable in the first stage (REL) is a binary variable, we follow the same methodology as that of Faulkender and Petersen (2006). The first stage probit is used to estimate the predicted probability of forming a relationship and this predicted probability is used as an instrument in the second stage estimation. Woolridge (2002) shows that this approach yields consistent coefficients as well as correct standard errors. Panel B of Table 5 also provides the results of our IV regression. The first stage F-statistic is 128.18 and rejects the null that the coefficients on the instruments are insignificantly different from zero, at the one percent level. We reproduce the OLS estimates (from Table 4, column 4 in the paper) and the IV results in Panel B of Table 5. The coefficient of relationship is -56.90 (significant at the five percent level). Thus, the effect of relationship on loan spreads is about five times higher when it is instrumented using log(1+distance). Berger, Miller, Petersen, Rajan, and Stein (2005) also use IV regressions to examine the impact of bank size on exclusivity of bank-borrower relationships. Using instruments for bank size they also find a large increase in coefficient for bank size compared to OLS 21 Shorter distance from the borrower increases the likelihood that lender and borrower would match up in the first place leading to relationship formation. Lower costs of screening and monitoring a physically close borrower is probably the key driver of relationships formation, which in turn affects the spreads charged. 22 Both Coval and Moskovitz (2001) and Dass and Massa (2006) provide details of the estimating formula. We use the same methodology. 23 estimates (approximately 5.5 times). The results of our IV regression suggest that relationships are associated with significant reduction in spreads for borrowers. 4.3.3 Treatment Effects Model The unverifiable assumption in an IV estimation is that the instrument is uncorrelated with the error term in the outcome equation. Empirically, there is no way to prove that the instrument is not correlated with the error term, since the error is by definition unobservable. To assess the robustness of our conclusions from our IV tests, we employ an additional empirical strategy that involves estimating the effect of an endogenously chosen binary variable (relationship or non-relationship) on another endogenous variable which is continuous (AISD), conditional on two sets of independent variables. Our set-up can also be addressed using the treatment effects model as described in Maddala (1983) and Greene (2000). We can write this as; AISDi = xi β + δRELi + ǫi (5) Since RELi is the endogenous variable, the binary decision to borrow from relationship lender is modeled as the outcome of an unobserved latent variable REL∗i which is assumed to be a linear function of exogenous covariates wi and a random component ui . Specifically, REL∗i = wi γ + ui (6) Thus the observed outcome is modeled as RELi = ( if REL∗i > 0 otherwise 1, 0, The key assumption is that u and ǫ are bivariate normal with mean zero and covariance matrix " # σ 2 ρσ ρσ 1 The treatment effects model differs from the instrumental variables estimation in terms of assumptions made about the error term. As explained above, the IV approach assumes that the instrument used is uncorrelated with the error term and this assumption cannot be verified for a model that is exactly identified. The treatment effects model on the other hand, assumes that the errors are bivariate normal - again this assumption cannot be empirically tested. Thus, both approaches have their strengths and weaknesses and we report results for both methodologies. We continue to use the log(1+distance) as before as an exogenous variable that affects the decision to 24 borrow from a relationship lender in the vector wi . This helps us identify the system and we do not simply depend upon non-linearities for identification. We report the results from the treatment effects model in Panel C of Table 5. The first column provides the estimated effect of distance on relationship. The coefficient is -0.055 (significant at the one percent level) implying, the greater the distance between borrower and lender, the lower is the likelihood of there being a relationship. The second column reports the endogeneity adjusted estimate of relationship on AISD. The coefficient is -16.91 (significant at the one percent level). The coefficient on the relationship variable is 1.5 times the size of the OLS estimate. Thus, after controlling for endogeneity, in a treatment effects model, we continue to find that past relationships are associated with significantly lower spreads. We also report the estimated correlation between the error terms in the two equations, ρ, which equals 0.042, which is positive and significant at the five percent level of significance. This signifies that any unobserved factor(s) not captured in our specification, that causes matching between a firm and a relationship lender and also affecting the observed spread, to be positively correlated. Of course, the treatment effects model is designed to explicitly take this correlation into account and provide the correct coefficients and standard errors. Thus, if unobserved credit quality causes both the formation/continuation of relationship as well as the observed spreads, the positive correlation implies that lower unobservable credit quality firms are more likely to borrow from relationship lenders. This result appears to alleviate a possible concern that banks might tend to build relationships with firms of unobservably higher credit quality (in which case ρ would have been negative). The central issue addressed in this section related to the potential endogeneity of the relationship variable itself. We addressed this issue in three different ways (PSM, IV estimation, and Treatment Effects Model). Our results continue to show that relationships matter for loan spreads using all three methods that account for endogeneity. The results are in fact somewhat stronger after addressing the endogeneity issue. 4.4 Multivariate Tests of H2 (Information Opacity) The multivariate tests in the section 4.2 suggest that there are significant benefits of relationships in terms of loan rates for our overall sample. The discussion in section 2.1.1 suggests that this overall benefit should vary systematically across different types of borrowers, with a greater level of benefit accruing to more informationally opaque borrowers. To test this, we propose the following model AISD = β0 + β1 (REL(M )) + β2 (Borrower Inf ormation Opacity) X + β3 (REL(M )) × (Borrower Inf ormation Opacity) + βi (Loan Charactersticsi ) X X βk (Controlk ). (7) + βj (Borrower Characatersticsj ) + 25 We employ six different measures of borrower information opacity in our tests: a) Log(Assets), b) a dummy variable for not rated firms, c) Inclusion in S&P 500 index, d) Number of analysts following the firm, e) A microstructure measure of information asymmetry, and f) level of discretionary accruals. The results are reported in Table 6. The coefficient for REL(Dummy)×Log(Assets) is 5.50 and is significant at the one percent level. The positive coefficient indicates that as firms become larger (less informationally opaque), the benefits of relationship lending declines, consistent with H2. Next, we use the existence of public bonds outstanding as a proxy for information opacity of the borrower. We create a dummy variable “Not Rated” which equals one if the firm does not have an S&P senior secured debt rating and zero otherwise. Faulkender and Petersen (2006) argue that the existence of an S&P debt rating is almost always associated with public debt outstanding.23 Arguably, unrated firms face higher information asymmetries since they are not monitored by credit rating agencies. For such firms, borrowing from relationship lenders should be especially beneficial. The results reported in the second column of Table 6 show that the interaction term REL(Dummy)× Not rated is -12.37 and is significant at the one percent level. This provides evidence in support of H2. We check also (in Column 3 of Table 6) to see if the borrower is included in the S&P 500 index at the time of loan. Firms that are part of this index are likely to be widely followed and can be considered informationally transparent. The interaction term is positive and significant implying that there is no significant benefit for such firms if they borrow from relationship lenders, again consistent with the view that it is the informationally opaque borrowers that benefit form relationship lending. We also match the number of analysts (column 4) that issue earnings forecasts for a particular borrower at the time of the loan. We obtain these data from I/B/E/S. Firms that have few or no analysts covering them can be thought of as informationally opaque firms where the benefits of relationship borrowing would be large. Column 4 provides evidence consistent with this interpretation, the interaction term is positive and significant at the one percent level, suggesting relationship benefits decline, as more analysts follow the firm. Columns 5 and 6 employ additional proxies of information opacity, a microstructure measure of information asymmetry (as described in Bharath, Pasquariello, and Wu, 2007) and the level of discretionary accruals (Dechow and Dichev, 2002). Higher levels for these measures imply a higher level of information asymmetry and correspondingly higher benefits from relationship borrowing and thus we expect the interaction term with REL(Dummy) to be negative. We find that the interaction terms are indeed negative and significant at the five and ten percent levels respectively. Taken together, the results reported in columns 1 through 6 of Table 6 provide strong empirical support for the hypothesis that repeated borrowing from the same lender 23 Cantillo and Wright (2000) report that fraction of firms that had a bond rating but did not have public debt was less than one percent. 26 is associated with larger reduction in spreads for firms with higher information opacity.24 These results are consistent with the view that relationships mitigate information asymmetry problems in contracting. Our earlier results show that for borrowers that break with their past lending relationship, the loans carry a higher spread (Table 5, Panel A). We also explore how break-ups interact with borrower opacity. We exclude all borrowers that never form relationships. This leaves us a with sample in which all borrowers that at some point had a lending relationship. We then create a dummy variable “Break” that takes the value one if the borrower had a past relationship but chooses to get the loan from a non-relationship lender (break of existing relationship) and zero otherwise. The discontinuation of existing relationship is likely to affect the opaque firms most acutely as the new lenders lack the proprietary information that would have been accumulated by the existing relationship lender and the gains of which were shared with the borrower. Our results are consistent with this argument. The interaction term (Opacity×Break) is significant and increasing in firm opacity for all the six different measures of borrower opacity.25 A potential criticism of our results is related to the concentration of the loan syndication market. Three banks, Citibank, JP Morgan Chase, and Bank of America, dominate the market for syndicated loans. Arguably, these banks can show up as lead banks on loans where they may not have a relationship with the borrower but may simply appear as repeated lenders due to their sheer market share. The last column in Table 6 controls for this effect by including dummy variable “Big-3 Bank” which equals 1 if any of the lead bank(s) is one of these three banks. The coefficient for REL(Dummy) is -14.39 (significant at the 1% level). These results imply that the lower rates on repeated borrowing from the same banks is not driven by the large market shares of the big three banks. We employ two additional robustness tests. First, we drop all loan facilities where any of the big three banks was involved in the role of the arranger. The coefficient for REL(Dummy) is -12.93 (significant at the one percent level). Second, we drop all facilities in which any of the big three banks was involved in any capacity. The coefficient for REL(Dummy) is -15.07 which is significant at the one percent level. Thus, our results are robust to the exclusion of big three banks.26 4.5 Multivariate Tests of H3 (Boundary of Relationship Lending) As discussed in section 2.1.1, Boot and Thakor (2000) posit that there should exist a critical point in terms of borrower quality below which banks engage in relationship lending and above which banks engage in transactional lending. Since several of our proxies for borrower informa24 The effect of lock in on spreads would result in exactly the opposite pattern: - Firms with higher information opacity should either have a lower cost saving or a higher spread relative to firms with lower information opacity. 25 Detailed results of these robustness tests are available from the authors on request. 26 Detailed results of these robustness tests are available from the authors on request. 27 tion opacity (e.g., size) are also correlated with observable borrower quality (e.g., existence of credit rating), the empirical results in the previous section provide fairly robust support for the conjecture. Our results show that the benefits of relationship lending decrease with a decrease in borrower opacity. To examine this in further detail, we revisit the results in Table 6 column 1 where size as measured by total assets is used as a proxy for borrower quality (higher the size, better is the borrower quality). The economic significance of this result is best illustrated by comparing the coefficient of the interaction term with the coefficient for REL(Dummy) which is -43.77. This specification allows us to measure the extent and significance of benefit of repeated borrowing from the same lender for different sized borrowers. Ranked by book value of assets, a borrower at the 10th percentile in our sample would have approximately 37 million in assets. This translates into a relationship benefit of roughly -24 bps (-43.77 +5.50×Log(37)). For the borrower at the 90th percentile level (asset size of 7,134 million) the net relationship coefficient is +5.07 bps (-43.77 +5.50×Log(7,134)), which is statistically insignificant. This exercise allows us to map the relative benefits of relationship loans across the borrower size spectrum. Figure 1 plots the coefficient of REL(Dummy) for different asset deciles and the t-statistic computed using the delta method for specification 1 in Table 6. The benefits of relationship decline as the borrower becomes larger and are statistically indistinguishable from zero for borrowers at the 70th percentile and above. This means for firms greater than $1.31 billion in asset size (in year 2000 dollars), relationship benefits as reflected in spreads disappear. We also use the Brown, Durbin, and Evans (1975) test to examine whether two regimes exist in our spreads regression with respect to borrower quality. Using the Brown-Durbin-Evans test for cumulative sum of square residuals, we find evidence that there are two regression regimes in our spreads regression and that the most likely point of the switch is around the 56th percentile in terms of borrower assets in year 2000 dollars (results not reported).27 Along similar lines, using the results in Table 27 Brown, Durbin, and Evans (1975) propose a method to evaluate the constancy of regression parameters for panel data sets. The null hypotheses is that there is no variation in all the regression parameters. The alternate hypothesis is that at least one of the regression parameters is not constant. Under the null hypothesis, the expected value of the regression residual (defined as a recursive residual) for the nth observation computed using parameter values estimated using the first n-1 observations is zero for all values of n. On the other hand, say the parameters switch at observation n. In that case, the expected value of the recursive residual for the nth observation using the parameter values computed using the first n-1 observation differs from zero. Using this insight, Brown, Durbin, and Evans (1975) suggest a test is called the ‘Cumulative Sum of Square Residuals’, that involves taking the sum of the square of the recursive residuals from sequential regressions consisting of the first ‘n’ observations. Brown, Durbin, and Evans show that this scaled sum of recursive residuals has a beta distribution and derive critical values for the hypothesis of constant regression parameters. Kallberg, Liu, and Srinivasan (2004) test for the most likely location of a switch point in the quality of limited partnerships using this method. In the context of testing the implication of the Boot and Thakor (2000) model, we are interested principally in constancy of the regression relationship over borrower quality. Thus, we are interested in evaluating the variation in regression parameters over borrower quality. To this end, we choose one of the main variables that we had earlier used to proxy for quality, namely size. We rank order the firms in our sample using asset size (in real 2000 dollars) and test for the constancy of regression parameters over variation in size. Figure 1 already provides strong and robust support for 28 6, column 2 where firm quality is measured by the presence of a credit rating, the coefficient for the interaction term (REL(Dummy)×Not Rated) is -12.37 (significant at the one percent level). Thus, on average, a relationship loan carries an interest rate 15 bps lower (obtained by summing the coefficients of relationship and interaction terms) for a firm that does not have public bonds outstanding compared to a similar borrower that has access to public debt markets. If ratings of public debt are used as a measure of borrower quality in a Boot and Thakor (2000) setting, our results imply that benefits of relationship dissipate significantly once a firm issues public debt. Similar interpretation can be extended to several of the other information opacity measures. For example, the benefits of relationship lending are insignificant if the borrower is part of S&P 500 index. The overall results of Table 6 along with Figure 1 provide support for the hypothesis that loans to firms above a certain quality threshold tend to be transactional loans while loans to firms of lower quality are relationship loans. Our results show that for such high-quality firms (proxied by borrower size) there is no significant difference in loan spreads based on whether the loan is from a relationship lender or a non-relationship lender. We describe such loans as “Transactional Loans” in this paper.28 4.6 Multivariate Tests of H4 (Syndicate Moral Hazard) The results so far are consistent with relationships addressing information asymmetry problems between lenders and borrowers. However, from the discussion in section 2.1.2, we know that relationships can also be viewed as a commitment to monitor, and further, this commitment will be especially valuable in syndicated loans where the lead bank has an incentive to reduce its monitoring effort by shirking. This syndicate moral hazard can exert an additional effect on the observed loan spreads. To isolate the impact of syndicate moral hazard we first focus on a subset of loans where this effect is not relevant. This sub-sample consists of loans where there was one reported lender. By construction, this sample lacks the moral hazard among lenders. The results reported in column 1 of Table 7, Panel A show that relationships are associated with almost 18 bps lowering of observed loan spreads (significant at the one percent level). When we further restrict this single lender sub-sample by requiring that the borrower has had a relationship with only one single lender through all of its prior loans, our results remain largely unchanged. The coefficient of REL is -14.44 which is significant at the five percent level (see column 2, Panel A, Table 7). The the non-constancy of the importance of the relationship on spreads, when sorted by firm size. However, the use of a formal statistical model of non-constant parameters can help us provide an independent method to verify that the benefits of relationships become insignificant as borrower size increases. 28 Simply because there is no observed reduction in loan spreads, does not imply by itself that these are transactional loans. After all, if there is no benefit to borrowing from a relationship bank why do it? We acknowledge that such loans may or may not be part of a larger “relationship bundle” where the benefits may appear in other products that the borrower may purchase from the lender. We leave tests of this issue for future research. 29 set of relationship loans in this sub-sample, where the borrower has had only one single bank relationship in the past and uses the same bank for the current loan, corresponds most closely to the theoretical scenario of “lock-in” envisaged in Sharpe (1990) and Rajan (1992). Even for this sub-sample, the relationship cost savings are comparable in magnitude to that obtained for the rest of the sample. The results in column 1 and column 2 indicate that relationships mitigate asymmetric information problems and that “lock-in” is not a significant issue for firms in our sample. Next, we try to estimate the impact of bank (or syndicate) moral hazard among the syndicate members on loan spreads. Sufi (2007) shows that syndicates tend to be more concentrated and the lead lender(s) retain(s) a significantly higher share of the loan when the likelihood of borrower moral hazard is high. Thus, increasing loan share retained by the lead bank and higher loan HHI imply higher moral hazard concerns. However, data on how the loan facility is shared among banks is available for less than half the sample. Therefore, in addition to the observed shares, we use the size of the syndicate as an additional proxy for syndicate moral hazard. Borrowers that require a high level of monitoring would be most affected by the syndicate moral hazard issue. One could argue that the observed size of a syndicate could be related to the perceived moral hazard by the syndicate members, with larger syndicates indicating lower moral hazard.29 The last three columns of Panel A of Table 7 provide empirical support for syndicate moral hazard playing a significant role in the determination of loan spreads. Larger syndicate size can be considered an ex-post indication of low moral hazard among syndicate members as seen in their willingness to join a large syndicate. We examine the role of relationship lending in the presence of syndicate moral hazard by including log of syndicate size and an interaction term of REL(Dummy) and log of syndicate size. If past relationships can be seen as credible commitments to monitor, such a commitment is likely to be less valuable where syndicate moral hazard is low (i.e., a large number of lenders are willing to join the syndicate). Overall, we find evidence consistent with this argument. Specifically, the coefficients on REL and the interaction term are -20.27 and 8.59 respectively (Column 3 of Table 7). The combined effect of relationship is estimated by summing the coefficients for REL(Dummy) and the interaction term. For the smallest syndicate size of one lender, the interaction term is zero. For such a borrower, having a past relationship implies a relationship benefit of 20.27 basis points. Since the median syndicate size is two for our sample, we consider the case of a loan that is provided by a syndicate that has the median size of our sample. Here the net effect of relationships is -14.31 basis points (-20.27+8.59×ln(2)). This result suggests that benefits of past relationships are lower when syndicate size increases (i.e., syndicate moral hazard is low). However, the overall effect of relationships at -14.31 basis points is still 29 As noted earlier in Section 2.1.2, there is no prior empirical work that has established a clear link between larger syndicate size and lower moral hazard. Therefore, the results on use of syndicate loan size should be seen as a secondary proxy for syndicate moral hazard, the primary ones being lead bank allocation and loan HHI. 30 statistically significant (p-value of 0.001). Sufi (2007) shows that high syndicate moral hazard is associated with larger loan allocation to the lead bank and higher Loan HHI. Again, we expect relationships to be of most value for loans where such moral hazard is high. The coefficients on REL and the interaction term between REL and lead bank allocation are 6.25 and -0.26 respectively (Column 4 of Table 7, Panel A). The combined economic and statistical effect of relationship is estimated by summing the coefficients for REL(dummy) and the interaction term. The median lead bank allocation in our sample is 68.75. Thus, for the median loan facility the total effect of a relationship is given by -11.86 basis points (6.25-0.26×68.75) which is significant at the one percent level (p-value of 0.001). Thus, past relationship provide significant benefits but those benefits are especially powerful for loans where lead bank retains a significantly large fraction since the interaction term is negative and significant. Consistent with this argument, assuming that the lead bank holds zero percent of the loan (suggesting little or no syndicate moral hazard) we should expect little benefits from relationships as a commitment to monitor. This is indeed what we find when we substitute lead bank allocation of zero percent into the specification in column 4 of table 7 - the interaction term drops out and the coefficient for REL (which is 6.75 basis points) is not significant. Column 5 uses the loan Herfindahl index as a measure of syndicate moral hazard. The coefficients for REL and REL×Loan HHI are 3.57 and -0.002. The median HHI in our sample is 6,250 implying a total relationship effect of -10.65 (3.57-0.002×6,250) which is significant at the one percent level (p-value = 0.003). Again, this implies an overall beneficial impact of relationship which is especially pronounced for loans with high HHI (high syndicate moral hazard). If Loan HHI is close to zero (implying little syndicate moral hazard), the interaction term drops out and as expected the coefficient for REL is not significant since relationships are not particularly prized when syndicate moral hazard is very low. We also test the relationship between observed loan spreads and syndicate moral hazard using propensity score matching (PSM). We need to divide our sample into high syndicate moral hazard and low syndicate moral hazard subgroups. For each loan where the entire loan is retained by the lead bank(s), we classify the given loan as one with low syndicate moral hazard, since by definition there is no syndicate moral hazard for this set of loans. For the remaining sub-sample, we calculate the median lead bank allocation. We classify loans with a lead bank allocation higher than the median as ones with high syndicate moral hazard, and loans with a lead bank allocation below the median as ones with low syndicate moral hazard. Using the PSM procedure, we create pairs of matched loans which are identical in all respects except that one loan in each pair would have a higher than median lead bank allocation, and the other loan in the same pair would have a lower than median lead bank allocation. Thus, for each pair, one loan (the one with high lead bank allocation) will have high syndicate moral hazard, while other loan (the one with low 31 lead bank allocation) will have low syndicate moral hazard. The observed difference in loan rates across all pairs (High - Low syndicate moral hazard) allows us to test the effect of syndicate moral hazard on loan spreads. We should expect this difference to be positive and significant. We use four matching methodologies as described earlier in the paper and we find a strong association between syndicate moral hazard and loan spreads, with high syndicate moral hazard loans having spreads of 6 to 10 bps higher than low syndicate moral hazard loans (significant at five percent level or lower).30 We also test if relationships play a meaningful role as a credible commitment to monitor. We focus on the subset of loans which are expected to have high syndicate moral hazard (i.e., loan syndicates where the lead bank allocation is higher than the sample median). For this group we use a probit model to estimate the probability of borrowing from a relationship bank (same model as Table 5 Panel A). The predicted probabilities allow us to create pairs of matched firms that are identical in all respects except for the relationship. One firm in each pair would have borrowed from its relationship lender while the other did not. The observed difference in loan rates across all pairs (Relationship - No Relationship) allows us to test the effect of relationships in a group that is pre-selected to have high moral hazard. If past relationships mitigate syndicate members’ concerns of shirking by the lead bank, the difference in spreads between relationship and non-relationship loans should be negative and significant. This is indeed what we find - for the high syndicate moral hazard group relationship loans are 11 to 19 bps lower compared to matched non-relationship loans and statistically significant.31 Finally, a typical loan provides a number of other contractual ways in which a lender can impact the “cost” of a loan to a borrower.32 These include extensive use of covenants and requirements for collateral. For example, a lender can cut off credit if a covenant is violated without having to go to extreme remedy of using foreclosure. By requiring collateral, a lender can commit to more rigorous monitoring.33 In Panel B of Table 7, we provide some limited evidence of past relationships being associated with a sharpening of effects of these contractual arrangements. In column 1 we interact REL with a dummy variable for Collateral. While a requirement of collateral is associated with higher spreads, the presence of past relationships 30 The four matching estimators are NEAREST NEIGHBOR (n=10), NEAREST NEIGHBOR (n=50), GAUSSIAN and EPANECHNIKOV methods. 31 These results are not reported but are available from the authors. 32 We employ covenant intensity and collateral as contractual features that necessitate intensive borrower monitoring (Rajan and Winton, 1995). We focus on the market to book ratio of borrowers as a measure of attractive growth opportunities. Arguably a lender can control a growing borrower (i.e., firm with a high market to book ratio) by cutting off credit if a covenant is violated without having to resort to foreclosure (which is expensive for both lender and borrower). For a low market to book firm, a lender may need to control such a firm by forcing a default, but this may impose costs on the lender as well. This illustrates that covenants and collateral can be especially useful contractual features for growing firms. Thus, relationships as a commitment to monitor, would make the loan especially valuable (hence lower spreads ex-ante) if covenants and collateral are present. 33 Collateral value may be sensitive to borrower actions, see Rajan and Winton (1995). 32 lowers the spread charged an average of 15 bps (significant at the 1% level). That is, collateral as a contractual device to mitigate borrower moral hazard is more effective in the presence of relationships, which serves as a commitment to monitor the borrower. We also construct a covenant index as described by Bradley and Roberts (2004). We focus on five specific covenants and assign one point to the index if that particular covenant condition is met. The covenants that we examine are: Asset Sweep, Debt Sweep, Equity Sweep, Dividend restriction, and presence of more than two financial ratio restrictions. Unlike Bradley and Roberts we exclude the secured covenant since we capture it through our collateral dummy. In column 2 we report how covenant intensity affects loan spreads when the borrower has a prior borrowing relationship with the lender through the interaction of REL(Dummy) with Covenant Index. Again, while a higher number of covenants by itself is associated with higher loan spreads, a past relationship in presence of a higher covenant index lowers the observed spread but the effect is not statistically significant. In the last two specifications (columns 3 and 4), we focus on borrowers that have attractive growth opportunities as evidenced by high market to book ratios. Such borrowers are more credibly controlled via threats of credit cut-offs in case of covenant violations. This is borne out by the negative and significant coefficient for the Market to book and covenant index interaction term. In column 4 we report how this benefit evolves if the firm has a past relationship with its lender. While the coefficient on the three-way interaction term is negative, it is not significant. 4.7 Multivariate Tests of H5 (Collateral) Next we focus on two specific non-price terms of loan contracts: collateral requirements and loan maturity. First, we present our findings on how past relationships and collateral requirements are inter-related. As discussed in section 2.2, collateral can also serve as a credible signal of quality to address the adverse selection problem (Bester, 1985, and Besanko and Thakor, 1987). Additionally, requiring collateral is a contractual mechanism to control borrower’s moral hazard incentives. If past relationships address the adverse selection problem and are credible commitments to monitor, then such relationship loans may require less collateral. To test this we run the following logit model: Collateral = β0 + β1 (REL(M )) + β2 Log(Loan Amount) + β3 (Leverage) + β4 (T angibility) + β5 (M arket to Book) + β6 (Loan Concentration) + β7 (Log(M aturity)) X + βk (Controlk ). 33 (8) Collateral is a dummy variable that equals 1 if the loan was secured and 0 otherwise.34 We control for loan amount, maturity, leverage, tangibility of assets, market-to-book, loan concentration (measured as the fraction of the loan size to the sum of existing debt plus the loan size), and default risk. We also include industry, facility purpose, and calendar year dummies. We Amount ) because if a particular loan facility is a significant use loan concentration ( ExistingLoan Debt+Loan Amount portion of the firm’s debt, it is more likely to be secured (Berger and Udell (1990); Boot, Thakor, and Udell (1991); Dennis, Nandy, and Sharpe (2000)).35 Our results are reported in Table 8. The three columns provide estimates for each of the REL measures. Regardless of which measure of relationship we employ, the coefficient is negative and significant at the one percent level. Thus, a loan from a relationship lender is significantly less likely to require collateral. These findings are consistent with the arguments of Holmstrom and Tirole (1997) that if a lending relationship is a credible commitment to monitor, such loans are less likely to require collateral. These results are also consistent with Bester (1985) that who suggests that the reduction in adverse selection by banking relationships would reduce collateral requirements. As reported in prior studies, we also find that loan facilities that are relatively large compared to existing debt are more likely to be granted on a secured basis. In particular, the coefficient for Loan Concentration is positive and significant at the one percent level. We also address the concern about the endogeneity of relationship and collateral. It is possible that an unobserved factor influences both the decision to form relationships and the requirement for collateral. We use two econometric methods to address the endogeneity of relationship in the collateral estimation. In particular, we use both an IV probit as well as bivariate probit methodologies. We use log(1+distance) as an instrument that determines relationships but not collateral. In the first stage, we use a probit model to estimate the probability of borrowing from a relationship lender. The predicted probability from this estimation is used as an instrument for the second stage probit that estimates the effect of relationship on collateral. We find that coefficient of relationship variable for the IV estimation is -1.11 and it is significant at the one percent level (for comparison, the coefficient on relationships in a probit equation (not correcting for endogeneity) for collateral generates a coefficient of -0.13 significant at the one percent level. Note that the results reported in Table 8 uses a logit specification where the coefficient for relationship variable is -0.22, significant at the one percent level). In the bivariate probit model, the coefficient for the relationship variable in the collateral equation is -1.33 and it is significant at the one percent level. These results of the IV probit as well as of the bivariate probit suggest 34 For nearly one-third of our loan sample LPC does not report whether the loans were secured by collateral or not. We treat such loans as unsecured. We also ran our tests by excluding all observations for which this data was missing. The results remain unchanged. 35 We also calculate a loan concentration ratio in which we exclude lines of credit and debt repayment and the results are unchanged. 34 that relationships are associated with lower collateral requirements, even after controlling for the endogeneity of the relationship variable.36 4.8 Multivariate Tests of H6 (Maturity) Next, we examine how past borrowing relationships affect (if at all) the maturity of the loan. As discussed in the theoretical section, the pure information asymmetry model by Flannery (1986) predicts relationships being associated with longer maturity debt while the Diamond (1991) model predicts that relationships could both increase or decrease debt maturity, depending on the credit rating of the firm. A factor complicating our estimation, especially in the context of the Diamond (1991) model, is that maturity is driven by borrower demand for high and intermediate credit rated firms and by lender supply for low rated firms. To get an insight into loan maturity-customer relationship effect we estimate several versions of the following basic model: Log(M aturity) = β0 + β1 (REL(M )) + β2 (Log(Loan Amount)) + β3 (Leverage) + β4 (Log(Assets)) + β5 (M arket to Book) + β6 (Log(Asset M aturity)) X + β7 (Regulated) + β8 (Collateral) + βk (Controlk ). (9) The dependant variable Log(Maturity) is the natural log of the stated maturity of the loan facility (measured as length in months between facility activation date and maturity date). We model the relationship between debt maturity and past banking relationships after controlling for variables that are known determinants for debt maturity (see Barclay and Smith (1995); Barclay, Marx, and Smith (2003)). We control for firm size, leverage, market-to-book, and two additional variables that are unique to the maturity regressions. We use a measure of asset maturity which is defined as the weighted average of maturity of current assets and Net PPE.37 The intuition behind this variable is that firms try to match their debt maturity to asset maturity. Hart and Moore (1994) describe a model of inalienability of human capital. This specificity of human capital makes the debt contracts hard to design since an entrepreneur can threaten to walk away from the project. They show that debt maturity can help address this problem. A key result of their model is that the optimal repayment path (debt maturity) would depend on the maturity structure of project payoffs and durability of assets. The model predicts that longer (shorter) life assets are likely to be financed by longer (shorter) term debt. Thus, using asset maturity as an explanatory variable for debt maturity is a reasonable assumption. As another factor that might explain debt maturity, we include a dummy variable for regulated 36 To conserve space we do not report these results but are available on request from the authors. We follow the definition provided by Barclay, Marx, and Smith (2003) to estimate Asset Maturity which equals CA CA NP P E NP P E (CA+N P P E) × COGS + (CA+N P P E) × Depreciation 37 35 industries following Barclay and Smith (1995). This is because the higher regulatory oversight for these firms should result in lower agency costs of debt, which, in turn, should result in greater use of longer maturity debt. Alternatively, if these regulated firms have access to longer maturity debt from capital markets, it might result in greater use of shorter maturity bank debt. In addition, we also control for default risk using the credit rating dummies and include year dummies and dummies for debt type and debt purpose. In one of the specifications, instead of the year dummies, we also include a term spread measure defined as the difference in yields on 10-year and 1-year U.S. Government bonds prevailing at the time of loan origination. Brick and Ravid (1985, 1991) provide a model in which the firm value is increasing if it issues long-term debt when the yield curve is upward sloping. Their argument is based on the fact that tax savings are accelerated when larger fraction of debt payment is allocated to long-term debt. This implies that maturity should be positively related to the term spread. The results of these regressions are presented in Table 9. The first column reports the coefficient on REL. We find that coefficient is negative and significant at the one percent level. However, the inference is clouded by the demand and supply effects identified by Diamond (1991). To test if there exists any cross-sectional variation of relationship effects on loan maturity across firm quality, we decompose the REL variable in three parts by interacting it with dummy variables High, Middle, and Low. High equals one if the borrower is ranked in the highest tercile as ranked by assets. Middle and Low are defined similarly. Interacting REL with each of these variables is equivalent to running separate regressions for each of the three size subgroups. Our results are reported in column 2. We find that effects of relationship are most pronounced for the highest and the lowest terciles (where the relationship coefficient is negative and significant), while the mid-sized borrowers’ debt maturity is not significantly affected by past relationships. In column 3, 4, and 5 we use the credit rating of the borrower as the measure of quality. Column 3 reports the maturity regression results for the subsample of firms that lack a credit rating. Presumably these firms can be viewed of being the lowest quality. Column 4 reports the same regression for firms that do have a credit rating but are below investment grade. These borrowers can be thought of as medium quality. Finally, in column 5 we tabulate maturity regression results for the sub-sample of investment grade-rated (high-quality) borrowers. Comparing the REL coefficient across these three groups shows a non-monotonic relationship between repeated borrowing from the same lender and credit quality. The worst and the best quality borrowers appear to have much shorter maturity loans from their relationship lender compared to medium-quality firms. We interpret these results as follows: For the highest-quality firms, relationship loans lower the cost of refinancing (liquidity risk) and therefore they demand even shorter loans from their relationship borrowers. For the lowest-quality firms, the relationship lenders commit to monitoring the firms more with shorter maturity loans. Given that monitoring costs for relationship lenders 36 are likely to be lower, we expect relationship loans to have a shorter maturity for low-quality borrowers. These results are consistent with Diamond (1991). The last column reports the association between maturity and the prevailing term spread. We find a significant and positive association as predicted by the models of Brick and Ravid (1985, 1991). 4.9 Instrumental Variables (IV) Estimation of Price and Non-Price Features of Loan Facilities Next, we address the issue of joint determination of price and non-price terms of loans. Previous empirical studies have largely ignored the potential interaction and trade-offs among the price and non-price terms of loans. To the extent that they are jointly determined, the true effects of relationships on these variables may be obscured. Melnik and Plaut (1986) model bank loans as a package of n contract terms that cannot be split and traded separately. Banks then offer an n-dimensional array of bundles from which to choose their contracts, and borrowers trade off contract terms in determining their optimal choice. This approach indicates that price, maturity, and collateral terms of a bank loan contract maybe interrelated. Consequently, we re-estimate the model specifications above for the price and non-price terms of loans using an IV framework. We describe the approach in detail below. As in Dennis et al. (2000) we assume a unidirectional relationship between the price (AISD) term and non-price (Collateral requirement and Maturity) terms. This assumed structure of joint determination needs some support. In particular, we are assuming that while maturity and collateral affect each other (bidirectional relationship), spread is only affected by maturity and collateral (unidirectional relationship). This structure reflects the loan syndication processes as described in discussions with industry professionals. For example, the S&P Guide to Loan Markets (2006) describes the process in several discrete steps. Syndication starts by the borrower appointing the lead bank, who conducts due diligence and hammers out the non-price loan features such as amount, collateral, maturity, and covenants with the borrower and leaves the final price (as yet) to be determined. At this stage, the lead bank would informally poll potential syndicate members to gauge the level of interest in the loan. This information is used to set the interest rate on the loan and it is launched for syndication. Thus, the loan syndication process has become increasingly similar to the “book-building” process used to sell publicly and privately placed bonds. This description of the syndication process makes our assumption of price being determined after all other non-price terms have been settled quite realistic. The three-equation structural model can be described as: 37 AISD = γA (REL) + γAC (Collateral) + γAM (Log(M aturity)) + XA βA + ǫA Collateral = γC (REL) + γCM (Log(M aturity)) + XC βC + ǫC Log(M aturity) = γM (REL) + γM C (Collateral) + XM βM + ǫM The γij are the coefficients of interdependence effects. Xk are the exogenous variables that affect the k th dependent variable. The estimation of the system is complicated by the fact that while spreads and maturity are continuous variables, collateral is a discrete choice variable. We therefore follow Woolridge (2002), who shows that estimating a logit equation for the discrete choice variable in the first stage and using the fitted value as an instrument for the discrete choice variable in the IV estimation, leads to consistent estimates of the coefficients. We employ the two-stage least squares (2SLS) method using instruments for our endogenous variables. Our choice of instruments is guided by existing theories of debt maturity, collateral, and price. We use prevailing default spread at the time the loan is made, as an instrument for observed spreads. Arguably, following the syndication process described by the industry professionals, contemporaneous default spreads should affect contracted loan spreads at the time of pricing (at which point, the other contract terms would have already been finalized). From the results reported earlier in Column 4 of Table 4 we also know that default spreads are positively and significantly related to loan spreads.38 We also created another instrument that determines loan rates (based on our conversations with bankers): the average AISD of loans completed over the previous six months. This average captures the recent evolution in loan pricing and is a significant factor in pricing of new loans. The Loan Pricing Corporation in its promotional materials for Dealscan highlights the use of these data in banks price-setting. These data are called Pricing Grids and are an important primary market pricing benchmark. These grids enable lenders/borrowers to gauge current spread levels and follow pricing trends by rating, size, and industry. Thus, it is reasonable to argue that the lagged average spread on loans over the last six months are related to the current loan’s AISD, but it is unlikely that past spread should affect this particular loan facility’s non price terms 38 For the default spread to be a valid instrument, one has to assume that changes in default spreads are uncorrelated with changes in collateral requirement. In order to investigate this issue, we constructed the proportion of loans in our database each month that were collateralized. We regress the contemporaneous fraction of loans that are collateralized on default spread, term spread, and GDP growth. The latter two determinants were included, following a) theoretical arguments by Kiyotaki and Moore (1997) that aggregate business cycles would determine the level of collateralization of loans and b) empirical arguments by Jiminez, Salas, and Saurina (2006). We estimate an OLS as well as generalized linear model (GLM). The default spread does not seem to be a determinant of the level of collateralization in both the OLS and GLM models. The coefficient on the default spread variable is insignificant and statistically indistinguishable from zero in all the specifications. Details of the test methodology and results are available on request from the author(s). 38 such as collateral requirements. In our spread regression we have two loan related endogenous variables, maturity and collateral, for which we need instruments. As described earlier, Hart and Moore (1994) argue that firms would attempt to match their debt maturity to the economic life of the assets. Barclay and Smith (1995) also use asset maturity as well as “regulated industry” dummy as key factors that affect the debt maturity structure of corporations. As an additional robustness check, following Brick and Ravid (1985, 1991) we also use the term spread as a factor that determines debt maturity. We use these variables as instruments for maturity. In addition loan concentration is employed as an instrumental variable that affects collateralization of the debt. The latter is based on the evidence in Berger and Udell (1990) who report that the greater the current loan borrowing relative to the size of the total debt, the greater is the likelihood of a lender asking for collateral. Following Berger and Udell (1990), we estimate Loan Concentration Amount . As an additional robustness test, we also use the industry median = (ExistingLoan Debt+Loan Amount) tangibility ratio as a factor that affects collateralization, with industries whose assets have greater tangibility more likely to provide collateral. The results of the IV estimation are presented in Table 10. Column 1 reports the OLS estimation for spreads (Column 4 of Table 4). Columns 2 and 3 report the effect of past relationships on loan spreads controlling for joint determination of maturity and collateral. The change of econometric specification produces some impact on the measured benefit of relationships on loan spreads. The coefficient for REL(Dummy) is -14.35 compared to -10.55 estimated in the simple OLS specification when we use default spreads as an instrument. Using average loan spreads over the previous 6 months yields a REL(Dummy) Coefficient of -12.79 again significant at the one percent level. Thus, once we control for endogeneity of various loan contract terms, our results show that borrowing from a relationship lender can reduce the cost of borrowing by up to 15 bps. These results are statistically significant at the five percent level. They are also economically significant as 15 bps represents an almost six percent savings on a median AISD of 212.5 bps. In the IV regression of Log(maturity), we find the effects of relationship are negative and significant for the low-quality firms and positive and significant for the medium-quality firms. Both of these results are consistent with Diamond (1991). While the coefficient for the highestquality borrowers with relationships is negative as predicted, it is not statistically significant. The coefficient of Collateral is positive and significant at the one percent level. This result confirms the prediction of Boot, Thakor, and Udell (1991) which implies that longer maturity loans are more likely to be collateralized.39 The coefficient for Asset Maturity, our exogenous variable for the maturity regression, is positive and significant at the one percent level, implying that firms try to match loan maturity to asset maturity. The coefficient for Regulated industry dummy, 39 Boot, Thakor, and Udell (1991) argue that “dissipative costs” i.e., likely loss in value of collateral, are higher for a longer maturity loan. Thus, they predict that longer maturity loans would be characterized by a higher level of collateral. 39 however, is negative; this is opposite of the results obtained by Barclay and Smith (1995). This difference is possibly due to the fact that we focus on bank loans while they examine public bond yields. Arguably, regulated firms borrow longer term from public debt markets and shorter term from banks. Given that Collateral is a dichotomous variable, we use an IV probit model. Again as expected, the probability of having to post collateral is significantly lower (at the one percent level) for borrowers who obtain loans from their relationship lenders. All other variables enter with predicted signs. Loan Concentration, our exogenous measure of collateral requirements, has a positive and significant coefficient suggesting that it is an appropriate instrument. In sum, the evidence presented in Table 10 suggests that our results as to the benefits of prior relationships are robust to simultaneous determination of contract terms. We also estimate the effects of relationship on price and non-priced terms using different instruments to assess its robustness. These include Term-Spread for maturity, median industry tangibility ratio for collateral, and interest coverage for spreads. The results 40 are very similar to the ones reported in Table 10. Overall, the relationship effect remains robust to these alternate IV estimations, confirming our single equation results. We also perform a variety of econometric tests to assess the relevance and validity of our instruments. These tests and their results are described and summarized in the next sub-section. 4.9.1 Tests of Relevance and Validity of Instruments We have used existing theoretical models and prior empirical research to motivate our instruments. However, our choice of instruments needs to be econometrically validated. The two key criteria that an instrument must meet are relevance (instrument is correlated with endogenous variable) and validity (instrument affects the dependent variable only through the endogenous variable). We conducted several tests that provide support for our instrument choices. The results of these tests are reported in the bottom Panel of Table 10. For our spread regression, we confirm that collateral and maturity are indeed endogenous by estimating the Durbin-Wu-Hausman (DWH) Chi-Squared test. The null hypothesis for this test is that the two variables, maturity and collateral, are exogenous to spread. Rejection of the null implies that these variables are indeed endogenous and validates the IV approach. We obtain a chi-squared test statistic of 469.25 when we use default spread as the instrument for observed spread which strongly rejects the null (pvalue = 0.000). Using average past spreads as an instrument we obtain chi-squared test statistic of 546.46 (p-value =0.000). To test if our instruments are relevant we conduct two tests to measure their strength. First we calculate the Cragg-Donald statistic which is 14.13 (for default spread as the instrument) and is higher than the 11.04 critical value reported by Stock and Yogo (2005) for 40 Available from the authors on request 40 an estimation with two endogenous variables. This implies that our instruments for maturity and collateral are relevant. Another test for relevance of instruments is the Anderson-LR test of the null hypothesis that correlations between the instrument and the endogenous variable is essentially zero. We obtain a test statistic value of 58.98 which strongly rejects the null (p-value =0.000), implying the instruments are strongly correlated with the endogenous variables. Results for using average spreads as instrument are very similar. Tests of the instrument’s independence from an unobservable error process (validity) can only be ascertained if the number of instruments is higher than the number of endogenous variables so that the system of equations are overidentified. Since we have two endogenous variables but four instruments for the spread equation, we are able to estimate the Hansen-J statistic for over-identification restrictions. These are joint tests of the null hypothesis that the correct model is specified and the orthogonality conditions are met (correlations between the instruments and the error term is zero). A rejection of the the null calls both of these assumptions into question. For our spread regression the HansenJ statistic is 4.13 with a p-value of 0.13. Thus, we fail to reject the null hypothesis, which implies that our instruments are relevant and valid. We repeat all of these tests described so far, for the specification in which we use the average AISD of the loans in the last 6 months as an instrument for AISD. These results are reported in column 3 of Table 10 and are similar to results obtained using default spread as the instrument. We report a Hansen J-statistic of 0.001 with a p-value of 0.98 for the maturity regression, which shows that our instruments for collateral in the maturity equation are valid. Since the maturity equation has only one endogenous variable (collateral), we report the first stage F-statistic and the Shea Partial R-Squared statistic instead of the Cragg-Donald statistic. We obtain a first-stage F-stat of 323.27 compared against the benchmark of 10 suggested by Staiger and Stock (1997) and a high partial R-Squared of 8.16%. These tests confirm the relevance of our instruments for collateral in the maturity equation. The Anderson LR statistic of 1289 with a p-value of 0.000 rejecting the null (of zero correlation between the instrument and endogenous variable) further strengthens our conclusions. Finally, the Durbin-Wu-Hausman (DWH) Chi-Squared test statistic of 1170 with a p value of 0.000 shows that collateral is indeed endogenous in the maturity equation. Since collateral is dummy variable we cannot apply the above tests of instrument relevance and validity.41 However, we do report the results of the Wald test to see if maturity is exogenous in the collateral equation. The null hypothesis that maturity is exogenous is strongly rejected (p-value =0.000). Collectively, these multiple tests provide econometric support for the relevance and validity of our instruments. 41 We are not aware of any econometric tests for IV probit models. 41 4.10 Multivariate Tests of H7 (Access to Financing) An additional potential benefit for a relationship borrower is increased access to financing. Faulkender and Petersen (2006) provide empirical evidence that even a number of publicly listed firms may be constrained in their ability to obtain debt financing. Furthermore, this constraint may be more binding for firms lacking strong borrowing relationships. Our univariate tests (Panel B, Table 3) provide partial evidence for this when we compared the leverage ratios of firms borrowing from their relationship lender versus those borrowing from non-relationship lenders. The relationship borrowers are significantly more leveraged than non-relationship borrowers. However, these univariate tests do not control for various other loan and borrower characteristics. To better understand how past relationships are associated with availability of bank loans we use multivariate tests. Since credit availability cannot be observed directly, we construct a proxy that attempts to measure the incremental loan amount available for borrowing. This is estimated as the ratio of loan amount being borrowed to the total assets of the borrower. To test if this ratio is related to the strength of relationships, we estimate a multivariate model of the following form. X Loan Amt = β0 + β1 (REL(M )) + βi (Loan Charactersticsi ) Assets X X βk (Controlk ). + βj (Borrower Charactersticsj ) + (10) Amt The dependent variable Loan is the ratio of dollar amount of a loan facility to the total Assets assets of the borrower. This ratio is the marginal amount of loan financing available to a borrower of a given size. A higher ratio implies better credit availability. The first three columns of Table 11 (Panel A) report the effect of prior relationships on credit availability. We use multiple specifications for the relationship variable: REL(Dummy), REL(Amount), and REL(Number). The coefficient for all specifications are positive and significant at the one percent level. The coefficient for REL(Dummy) is 0.008, which implies that on average, borrowing from a relationship lender would be associated with almost one percent larger loan availability (as a percent of assets) compared to borrowing from a non-relationship lender. In the last three columns of Table 11 we re-estimate our model using a different proxy for credit availability. We scale the loan amount by existing total long-term debt of the borrower. The results are essentially unchanged. For example, the coefficient for REL(Dummy) is 0.02, implying a two percent increase in loan availability. Thus, we find that not only are relationship loans marked by better spreads and lower collateral, they also tend to be associated with greater credit availability. 42 5 Conclusion The role of strong relationships between lenders and borrowers has been an active area of research. Theoretical models have generated a number of rich empirical implications both in terms of how past relationships may affect loan contracts in the future as well as a set of borrower characteristics that would make relationships more (or less) important. We find that repeated borrowing from the same lender translates into a 10 to 17 bps lowering of loan spreads. Our results provide evidence that hypothesized economies in information production due to repeated interaction between the same lender and borrower are at least partly reflected in the price of loans. This result continues to hold across various methodologies that control for the endogeneity of the relationship formation. We also find that as the information opacity of a borrower increases, the observed reduction in the cost of borrowing due to a relationship becomes greater. These results also shed light on the boundary of relationship and transactional lending. We find that spreads charged for relationship loans and non-relationship loans become indistinguishable if the borrower was in the top 30% when ranked by asset size. Similar dissipation of relationship benefits occurs if the borrower has a rated public debt or is part of S&P 500 index. The paper uses loan syndicate structure to examine the interaction of past relationships and syndicate moral hazard due to multiple lenders. The results in our paper indicate that past relationships can mitigate syndicate moral hazard issues by serving as a commitment to monitor. We also find that relationship loans are significantly less likely to be secured by collateral, and our results on loan maturity are broadly consistent with Diamond (1991). Our results are robust to an estimation methodology which allows loan spread, collateral requirement, and loan maturity to be determined jointly using an IV approach. Finally, the relationship borrowers also obtain larger loans (scaled by the borrower’s asset size or long-term debt) compared to non-relationship borrowers. In sum, we report significant benefits of borrowing from relationship lenders even for publicly traded firms. 43 APPENDIX Hidden Bias in Propensity Score Matching Results In order to estimate the extent to which such “selection on unobservables” may bias our qualitative and quantitative inferences about the effects of relationships on loan spreads, we present the results of a Rosenbaum bounds sensitivity analysis (Rosenbaum, 2002). The basic question is whether unobserved factors can alter inference about treatment effects (the effect of relationships on spreads). One wants to determine how strongly an unmeasured variable must influence the selection process of forming relationships to undermine the implications of the matching analysis. The bounding approach does not test the unconfoundedness assumption itself, because this would amount to testing that there are no (unobserved) variables that influence the selection into treatment. Instead, Rosenbaum bounds provide evidence on the degree to which any significance results hinge on this untestable assumption. Let us assume that the relationships participation probability is given by Pi = P (xi , ui ) = P (RELi = 1|xi , ui ) = F (β xi + γ ui ) (A-1) where xi are the observed characteristics for borrower i, ui is the unobserved variable, and γ is the effect of ui on the participation decision. If our study is free of hidden bias, γ will be zero and the participation probability will be determined solely by xi . However, if there is hidden bias, two borrowers with the same observed covariates x have different chances of receiving treatment. Let us assume that we have a matched pair of borrowers i and j and further assume that F is the logistic distribution. The odds that borrowers Pj Pi borrow from a relationship lender (receive treatment) are then given by (1−P and (1−P , and the odds ratio is given by ) ) i j Pi (1 − Pj ) exp(βxi + γui ) = Pj (1 − Pi ) exp(βxj + γuj ) (A-2) If both borrowers have identical observed covariates as implied by the matching procedure, the x vector cancels out, implying that Pi (1 − Pj ) = exp(γ(ui − uj )) Pj (1 − Pi ) (A-3) But still, both borrowers differ in their odds of receiving treatment by a factor that involves the parameter γ and the difference in their unobserved covariates u. So, if there are either no differences in unobserved variables (ui = uj ) or if unobserved variables have no influence on the probability of participating (γ = 0), the odds ratio is one, implying the absence of hidden or unobserved selection bias. Sensitivity analysis now evaluates how changing the values of γ and (ui , uj ) alters inference about the treatment effect. We follow Rosenbaum (2002) and assume for simplicity that the unobserved covariate is a dummy variable with u ∈ 0, 1. He shows that the above discussion implies the following bounds on the odds ratio that either of the two matched borrowers will receive treatment: 1 exp(γ) ≤ Pi (1 − Pj ) Pj (1 − Pi ) ≤ exp(γ) (A-4) Both matched borrowers have the same probability of participating only if exp(γ) = 1. Otherwise, if for example exp(γ) = 2, borrowers who appear to be similar (in terms of x) could differ in their odds of receiving the treatment by as much as a factor of 2. In this sense, exp(γ) is a measure of the degree of departure from a study that is free of hidden bias. Further, Rosenbaum (2002) develops a test statistic T (a Wilcoxon signed rank test statistic) for matched pairs where the outcome for the treatment is greater than the outcome for control. The ranks of these cases are summed and compared with the distribution of the test statistic under the null hypothesis that the treatment has no effect. T = t(Z, r) = »X S s=1 ds 2 X i=1 csi Zsi – (A-5) where Z is the variable that records which of each of the s pairs was treated, and r is the outcome for each case in the S pairs. Zsi equals one if a case is treated, and 0 otherwise; c is defined as follows: 8 < cs1 = 1, cs2 = 0, cs1 = 0, cs2 = 1, c: : cs1 = 0, cs2 = 0, if rs1 > rs2 if rs1 < rs2 if rs1 = rs2 Finally, ds is the rank of |rs1 − rs2 | with average ranks used for ties. As Rosenbaum shows in the case where the assignment to the treatment is not random, the above test statistic can be bounded. Under the assumption that a confounding variable u exists, the formula for T is the sum of S independent random variables where the sth pair equals ds with probability ps = cs1 exp(γ us1 ) + cs2 exp(γ us2 ) exp(γ us1 ) + exp(γ us2 ) and equals zero with probability 1-ps . Define 44 (A-6) p+ s : 8 < 0, : and p− s : if cs1 = cs2 = 0 exp(γ) , 1 + exp(γ) if cs1 6= cs2 8 < 0, if cs1 = cs2 = 0 1 , : 1 + exp(γ) if cs1 6= cs2 Rosenbaum (2002) shows that for a given value of γ, the null distribution of T = t(Z, r) is bounded by two known distributions for T + and T − , where E(T + ) = E(T − ) = S X s=1 S X ds p+ s , ds p− s , s=1 V ar(T + ) = V ar(T − ) = S X s=1 S X s=1 + d2s p+ s (1 − ps ), − d2s p− s (1 − ps ) We can use these formulas to compute the significance level of the null hypothesis of no effect. For any specific γ, we compute (T −E(T + )) √ V ar(T + ) (T −E(T , √ − )) V ar(T − ) where T is the Wilcoxon signed rank statistic. These two values give the bounds of the significance level (p-values) of a one sided test for no effect of the treatment (i.e., relationships). The table below reports p-values from Wilcoxon signed rank tests for the average treatment effect (effect of relationship on spreads) on the treated while setting the level of hidden bias to a certain value exp(γ). Value of exp(γ) reflects our assumption about endogeneity in treatment assignment in terms of the odds ratio of differential treatment assignment due to an unobserved covariate (equation A-3). At each exp(γ), we calculate a hypothetical significance level “p-critical” (both a lower bound and an upper bound, corresponding to the bounds on the odds ratio), which represents the bound on the significance level of the treatment effect in the case of endogenous selection into treatment status. The method for calculating these variables has been outlined above. We do this for the matched pairs obtained using the nearest neighbor (n=10) matching method. The results on the bounds are similar when calculated for other matching methods. By comparing the Rosenbaum bounds on treatment effects at different levels of exp(γ) (the recommendation is to increase exp(γ) in steps of 0.05, starting from 1), we can assess the strength such unmeasured influences must have in order that the estimated treatment effects from propensity score matching would have arisen purely through non random assignment.Intuitive interpretation of exp(γ) being equal to one is that unobservable factor(s) (hidden bias) has no effect in how each individual record (in our case each loan facility) is assigned to a treatment group (e.g., being classified as a relationship loan) or to a control group (e.g., being classified as a non-relationship loan). For values of exp(γ) higher than one, implication is that the hidden bias is increasing the likelihood of being assigned to one group compared to being assigned to the other group. exp(γ) 1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.50 Rosenbaum Bounds p-critical Effect on Spreads upper bound lower bound upper bound lower bound 0 0 -22.88 -19.06 0 0 -25.25 -16.63 0 0 -27.50 -14.29 0 0 -29.63 -12.05 0 0 -31.69 -9.88 0 6.70E-16 -33.63 -7.81 0 1.20E-10 -35.50 -5.79 0 9.80E-07 -37.25 -3.88 0 0.000609 -38.94 -2.00 0 0.038583 -40.56 -0.13 0 0.363088 -42.13 1.63 45 Under the more stringent assumption of constant treatment effect, Rosenbaum (2002) also calculates the bounds on the point estimate of the treatment effect. This enables the researcher to frame the sensitivity analysis in terms of the interval of point estimates at a given confidence level. These results are shown in the Table above, under the column titled, “Effect on Spreads.” The critical level of exp(γ) at which we would have to question our conclusion of a negative effect of relationships on spreads is 1.50 (it is at this level that the confidence interval contains zero for the first time). A exp(γ) of 1.50 is attained if an unobserved covariate caused the odds ratio of treatment assignment to differ between treatment and control cases by a factor of about 1.50. It is important to recognize that these results are worst-case scenarios. A value for exp(γ) of 1.50 does not mean that there is no true negative effect of relationships on spreads. This result means that the confidence interval for the relationship effect on spreads includes zero, (a) If an unobserved variable caused the odds ratio of treatment assignment to differ between the treatment and control groups by 1.5, and (b)The unobserved variable’s effect on outcome (spreads) has to be so strong as to almost perfectly determine whether the observed outcome (spread) would be bigger for the treatment case (relationship) in each pair of matched cases in the data, so as to overturn our inference and include zero in the confidence interval for the estimated effect. Note, in the case where a confounding variable had an equally strong effect on group assignment (treatment versus control) but only a weak effect on the outcome variable, the confidence interval for spreads would not contain zero. To illustrate the magnitude of hidden bias that would cause us to revise our findings of causal effects of relationships on spreads, we equate the magnitude of hidden bias expressed by the specific level of exp(γ) = 1.50 in terms of the equivalent effect of observed covariates for which we know the impact on assignment to treatment from our propensity score model. From the logistic regression model (with a covariate xk with coefficient βk and standard deviation sk ) , for a n standard deviation in variable xk we know that the odds ratio are expected to change by a factor of exp(βk ∗ sk ∗ n), holding all other variables constant. By setting this equal to a factor of 1.5 (the change in odds ratio from exp(γ) = 1.5 from exp(γ) = 1), we can assess n for each observed covariate in the model. We do this for the covariates that are commonly thought to control for credit risk. This exercise yields the table below: n standard deviation in observed covariate required to justify an increase in odds ratio by a factor of 1.5 holding all other variables constant in the propensity score model Variable Odds Ratio Std.Dev Implied exp(βk ) sk n Assets 0.987 1.872 -16.55 Profitability 0.899 0.120 -31.73 Log (1+Coverage) 1.189 1.106 2.12 Leverage 2.194 0.189 2.73 Default Spread 1.117 0.582 6.30 The above table indicates that implausibly high changes in observed credit risk covariates are necessary to increase the odds ratio by a factor of 1.5. For example, leverage will have to increase by at least 50%, (i.e.) 2.73 standard deviations to increase the odds ratio of assignment by a factor of 1.5, which is the point at which we begin to question our results. Based on these results for observed covariates that affect credit risk, we conclude that it is unlikely such powerful unobserved covariates (over and above the long vector of observed covariates that we have controlled for) can be at work to challenge our estimates of the causal effect of relationships on spreads. 46 References [1] Barclay, Michael J., and Clifford W. Smith, 1995, The Maturity Structure of Corporate Debt, Journal of Finance, 50, 609-631. [2] Barclay, Michael J., L.M. Marx and Clifford W. Smith, 2003, The Joint Determination of Leverage and Maturity, Journal of Corporate Finance, 9, 149-167. [3] Berger, Allen, and Gregory Udell, 1990, Collateral, Loan Quality and Bank Risk, Journal of Monetary Economics, 25, 21-42. [4] Berger, Allen, and Gregory Udell, 1995, Relationship Lending and Lines of Credit in Small Firm Finance, Journal of Business, 68, 351-381. [5] Berger, Allen, Marco Espinosa-Vega, Scott Frame, and Nathan Miller, 2005, Debt Maturity, Risk, and Asymmetric Information, Journal of Finance, 60, 2895-2923. [6] Berger, Allen, N. Miller, M. Petersen, R. Rajan, and J. Stein, 2005, Does function follow organizational form? Evidence from the lending practices of large and small banks, Journal of Financial Economics, 76, 237-269. [7] Besanko, David, and Anjan Thakor, 1987, Collateral and Rationing: Sorting Equilibria in Monopolistic and Competitive Credit Markets, International Economic Review, 28, 601-689. [8] Bester, Helmut, 1985, Screening vs. Rationing in Credit Market under Asymmetric Information, Journal of Economic Theory, 42, 167-182. [9] Berlin, Mitchell, and Loretta Mester, 1992, Debt Covenants and Renegotiations, Journal of Financial Intermediation, 2, 95-133. [10] Berlin, Mitchell, and Loretta Mester, 1998, Deposits and Relationship Lending, Review of Financial Studies, 2, 95-133. [11] Bharath, Sreedhar, Sandeep Dahiya, Anthony Saunders, and Anand Srinivasan, 2007, So What Do I Get? The Bank’s View of Lending Relationships, Journal of Financial Economics, 85, 368-419. [12] Bharath, Sreedhar, Paolo Pasquareillo, and Guojun Wu, 2007, Does Information Asymmetry Drive Capital Structure Decisions, Forthcoming, Review of Financial Studies. [13] Bhattacharya, Sudipto, and Gabriella Chiesa, 1995, Proprietary Information, Financial Intermediation and Research Incentives, Journal of Financial Intermediation, 4, 328-357. [14] Boot, Arnoud, 2000, Relationship Banking: What Do We Know?, Journal of Financial Intermediation, 9, 7-25. [15] Boot, Arnoud, Anjan Thakor, and Gregory Udell, 1991, Secured Lending and Default Risk: Equilibrium Analysis, Policy Implications and Empirical Results, The Economics Journal, 101 (406), 458-472. [16] Boot, Arnoud, Stuart Greenbaum, and Anjan Thakor, 1993, Reputation and Discretion in Financial Contracting, American Economic Review, 83, 1165-1183. [17] Boot, Arnoud, and Anjan Thakor, 1994, Moral Hazard and Secured Lending in an Infinitely Repeated Credit Market Game, International Economic Review, 35, 899-920. [18] Boot, Arnoud, and Anjan Thakor, 2000, Can Relationship Banking Survive Competition?, Journal of Finance, 55, 679-713. [19] Bradley, Michael and M.Roberts, 2004, The Structure and Pricing of Corporate Debt Covenants. Available at SSRN: http://ssrn.com/abstract=466240 [20] Brick, Ivan, and Abraham Ravid, 1985, On the Relevance of Debt Maturity Structure, Journal of Finance, 40, 1423-1437. 47 [21] Brick, Ivan, and Abraham Ravid, 1991, Interest Rate Uncertainty and the Optimal Debt Maturity Structure, Journal of Financial and Quantitative Analysis, 26, 363-381. [22] Brown, R. L., J. Durbin, and J. M. Evans, 1975, Techniques for Testing for the Constancy of Regression Relationships over Time, Journal of the Royal Statistical Society, 37, 149-192. [23] Cantillo Miguel and Julian Wright, 2000, How Do Firms Choose Their Lenders? An Empirical Investigation, Review of Financial Studies, 13, 155-189. [24] Cole, Rebel A., 1998, The Importance of Relationships to the Availability of Credit, Journal of Banking & Finance, 22, 959-977. [25] Coval, Joshua D. and Moskowitz, Tobias J., 2001, The Geography of Investment: Informed Trading and Asset Prices, Journal of Political Economy, 109 (4), 811-841. [26] Dahiya, Sandeep, Anthony Saunders, and Anand Srinivasan, 2003, Financial Distress and Bank Lending Relationships, Journal of Finance, 58, 375-399. [27] Dass, Nishant and Massimo Massa, 2006, The Bank-Firm Relationship: A Trade-off Between Better Governance and Greater Information Asymmetry, Working Paper, INSEAD. [28] Dechow, Patricia, and I. Dichev, 2002, The Quality of Accruals and Earnings: The Role of Accrual Estimation Errors, The Accounting Review, 77. Supplement, pp. 35-59. [29] Degryse, Hans, and Patrick Van Cayseele, 2000, Relationship Lending within a Bank Based System: Evidence from European Small Business Data, Journal of Financial Intermediation, 9, 90-109. [30] Dennis, Steven, Debarshi Nandy, and Ian Sharpe, 2000, Determinants of Contract Terms in Bank Revolving Lines of Credit, Journal of Financial and Quantitative Analysis, 35, 87-109. [31] Detragiache, Enrica, Paolo Garella, and Luigi Guiso, 2000, Multiple versus Single Banking Relationships: Theory and Evidence, Journal of Finance, 55, 1005-1476. [32] Diamond, Douglas, 1984, Financial Intermediation and Delegated Monitoring, Review of Economic Studies, 62, 393-414. [33] Diamond, Douglas, 1991, Debt Maturity and Liquidity Risk, Quraterly Journal of Economics106(3), 709-737. [34] Diamond, Douglas, 2004, Presidential Address, Committing to Commit: Short Term Debt When Enforcement Is Costly, Journal of Finance, 59, 1447-1479. [35] Drucker, Steven, and Manju Puri, 2005, On the Benefits of Concurrent Lending and Underwriting, Journal of Finance, 60, 2763-2799. [36] Fama, Eugene., 1985, What’s Different About Banks? Journal of Monetary Economics 15(1), 29-39. [37] Faulkender. Michael, and Mitchell Petersen, 2006, Does the Source of Capital Affect Capital Structure? Review of Financial Studies, 19(1), 45 - 79. [38] Flannery, Mark, 1986, Asymmetric Information and Risky Debt Maturity Choice, Journal of Finance, 41, 19-37. [39] Greene, W. H., 2000, Econometric Analysis, 4th ed, Upper Saddle River, NJ: PrenticeHall. [40] Hart, Oliver, and John Moore, 1994, A Theory of Debt Based on the Inalienability of Human Capital, Quarterly Journal of Economics, 109, 841-79. [41] Heckman, James, Hidehiko Ichimura, and Petra Todd, 1997, Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme, Review of Economic Studies 64, 605-654. [42] Heckman, James, Hidehiko Ichimura, and Petra Todd, 1998, Matching as an Econometric Evaluation Estimator, Review of Economic Studies 65, 261-294. 48 [43] Holmstrom, Bengt and Jean Tirole, 1997, Financial Intermediation, Loanable Funds and the Real Sector, Quarterly Journal of Economics, 112(3), 663-691. [44] Hoshi, Takeo, Anil Kashyap, and David Scharfstein, 1990, The Role of Banks in Reducing the Costs of Financial Distress in Japan, Journal of Financial Economics, 27, 67-88. [45] Jiménez, Gabriel, Vicente Salas, Jesús Saurina, 2006, Determinants of Collateral, Journal of Financial Economics, 81(2), 255-281. [46] Kallberg. J, C. Liu, and A. Srinivasan, 2004, Limited Partnerships and Reputation Formation, Journal of Financial and Quantitative Analysis, 39, 631-659. [47] Kiyotaki, N., and J. Moore, 1997, Credit Cycles Journal of Political Economy, 105, 211 - 48. [48] Melnik, Arie and Steven Plaut, 1986, Loan Commitment Contracts, Terms of Lending, and Credit Allocation, Journal of Finance, 41, 425-435. [49] Maddala, G. S., 1983, Limited-dependent and Qualitative Variables in Econometrics, Cambridge, UK: Cambridge University Press. [50] Petersen, Mitchell, and Raghuram Rajan, 1994, The Benefits of Lending Relationships: Evidence from Small Business Data, Journal of Finance, 49, 3-37. [51] Petersen, Mitchell, and Raghuram Rajan, 1995, The Effect of Credit Market Competition on Lending Relationships, Quarterly Journal of Economics, 110, 406-443. [52] Petersen, Mitchell A., and Raghuram G. Rajan, 2002, Does Distance Still Matter? The Information Revolution in Small Business Lending, Journal of Finance, 57, 2533- 2570. [53] Rajan, Raghuram, 1992, Insiders and Outsiders: The Relationship Between Relationship and Arms Length Debt, Journal of Finance, 47, 1367-1400. [54] Rajan, Raghuram, and Andrew Winton, 1995, Covenants and Collateral as Incentives to Monitor, Journal of Finance, 50,1113-1146. [55] Ramakrishnan Ram, and Anjan Thakor, 1984, Information Reliability and a Theory of Financial Intermediation, Review of Economic Studies, 51, 415-432. [56] Rogers, William, 1993, Regression Standard Errors in Clustered Samples, Stata Technical Bulletin 13, 19-23. [57] Rosenbaum, Paul R., 2002, Observational Studies, 2nd Edition, Springer-Verlag, New York. [58] Sharpe, Steven, 1990, Asymmetric Information, Bank Lending and Implicit Contracts: A Stylized Model of Customer Relationships, Journal of Finance 45, 1069-1087. [59] Standard & Poor’s, 2006, A Guide to the Loan Market. [60] Staiger, Douglas and James, Stock, 1997, Instrumental Varaiables Regression with Weak Instruments, Econometrica, 557-586. [61] Stock James, and Motohiro, 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 J. Rothenberg. Cambridge: Cambridge University Press. 80-108. [62] Strahan, Phillip, 1999, Borrower Risk and the Price and Non-price Terms of Bank Loans, Federal Reserve Bank of New York, Staff Paper No. 90, Available at SSRN: http://ssrn.com/abstract=192769 [63] Stulz, René, and Herb Johnson, 1985, An Analysis of Secured Debt, Journal of Financial Economics, 14, 501-521. [64] Sufi, Amir, 2007, Information Asymmetry and Financing Arrangements: Evidence from Syndicated Loans Journal of Finance, 62, 629-668. [65] Wooldridge, J. M., 2002, Econometric Analysis of Cross Section and Panel Data, The MIT Press, Cambridge, Massachusetts. 49 TABLE 1 Descriptive Statistics of Loan Facilities The table below provides the descriptive statistics for the sample of loan facilities. The statistics are reported separately for loans from relationship lenders REL(Dummy) =1 and loans from non-relationship lenders REL(Dummy)=0. For any particular loan facility REL(Dummy) equals one if any of the lead lenders for that loan facility had been a lead lender on any loans to that borrower in the 5 years preceding the loan facility. Panel A :Calendar Time Distribution of Loans Year of Loan No Relationship Relationship Sanction REL(Dummy) = 0 REL(Dummy) = 1 Total 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 Total 3 116 428 569 631 723 926 1,149 1,345 1,377 1,708 2,115 1,975 1,820 1,943 1,744 1,583 1,477 21,632 1 68 217 252 242 276 363 412 421 359 598 618 609 548 519 474 448 375 6,800 2 48 211 317 389 447 563 737 924 1,018 1,110 1,497 1,366 1,272 1,424 1,270 1,135 1,102 14,832 50 TABLE 1(Continued) Panel B: Industry Classification of Borrowers One Digit No Relationship Relationship SIC Code REL(Dummy) = 0 REL(Dummy) = 1 0 23 54 1 372 915 2 950 2,210 3 1,706 3,357 4 682 1,820 5 895 1,928 7 602 1,360 8 229 539 9 31 40 Total 5,490 12,223 Panel C: Loan Purpose Year of Loan No Relationship Sanction REL(Dummy) = 0 Acquisition line 268 Corporate purposes 1,973 CP backup 190 Debt Repayment 1,599 Debtor-in-possession 98 LBO/MBO 339 Recapitalization 165 Takeover 735 Working Capital 1,066 Other 366 Total 6,799 51 Total 77 1,287 3,160 5,063 2,502 2,823 1,962 768 71 17,713 Relationship REL(Dummy) = 1 646 3,974 1,483 3,906 197 304 297 1,621 1,813 590 14,831 Total 914 5,947 1,673 5,505 295 643 462 2,356 2,879 956 21,630 TABLE 2 Summary Statistics for Key Loan and Borrower Characteristics The table below provides summary statistics of various loan and borrower characteristics. Panel A reports key characteristics of the Loan Facilities and Panel B reports the same for borrower characteristics. AISD is the “All In Spread-Drawn,”, which is the all-inclusive cost of a drawn loan to the borrower. This equals the coupon spread over LIBOR on the drawn amount plus the annual fee and is reported in basis points. Loan Facility Size is the dollar amount of loan facility in millions. AISU is the fee charged on the undrawn loan amounts. AISD, AISU, and Annual Fee are reported in basis points. Maturity is length in months between facility activation date and maturity date. Syndicate, Collateral, and Investment Grade are percent of facilities that have the stated attribute. To be classified as Investment Grade the loan has to be rated BBB or above by S&P and zero otherwise. The descriptive statistics for Investment Grade are estimated for the subset of firms that are rated. Asset is the book value of assets of the borrower as reported in the COMPUSTAT. Leverage is the ratio of book value of total debt to book value of total assets. Coverage is the ratio of EBITDA to interest expenses. Profitability is the ratio of EBITDA to Sales. Tangibility is the ratio of NPPE to Total Assets. Current Ratio is the ratio of Current Assets to Current Liabilities. Market to book is the ratio of (Book value of assets - Book value of equity + market value of equity) divided by book value of assets. All values are winsorized at the 1% and 99% level. Variable N Mean Std. Dev. Min 25th Pctile Median 75th Pctile Max 26,955 15,265 8,969 6,634 31,384 28,335 19,813 31,387 216.63 33.38 66.37 19.52 190.0 43.2 0.82 0.79 139.89 23.45 72.94 25.63 500.0 28.36 0.39 0.41 17.50 4.00 2.50 1.33 0.5 3 0.0 0.0 100.00 15.00 20.00 7.50 13.0 17 1.0 1.0 212.50 30.00 50.00 12.50 50.0 36 1.0 1.0 300.00 50.00 100.00 24.50 180.0 60 1.0 1.0 630.00 125.00 315.00 100.00 2000.0 120 1.0 1.0 2,941.4 40.18 0.25 0.14 0.34 2.08 1.75 0.50 0.73 11,279.0 731.51 0.19 0.13 0.23 2.50 1.45 0.50 0.45 6.6 0 0 0 0.02 0.32 0.66 0.0 0.0 93.3 2.4 0.08 0.06 0.16 1.14 1.09 0.0 0.0 360.4 4.83 0.23 0.12 0.29 1.66 1.36 0.0 1.0 1,623.5 10.05 0.37 0.19 0.50 2.41 1.90 1.0 1.0 38,750.0 309.40 0.82 0.62 0.90 8.86 7.22 1.0 1.0 Panel A: Loan Characteristics 52 AISD (Basis Points) AISU (Basis Points) Fee Upfront (Basis Points) Fees Annual (Basis Points) Loan Facility Amount ($ Millions) Maturity of Loan (Months) Collateral Syndicate Panel B: Borrower Characteristics Borrower Assets ($ Millions) Coverage Leverage Profitability Tangibility Current Ratio Market to Book Investment Grade Not Rated 23,679 22,604 23,490 23,838 23,596 22,578 23,463 8,590 31,387 TABLE 3 Key Price and Non-Price Loan Contract Terms - Relationship versus Non-Relationship Loans Panel A segregates the entire sample in non-relationship (REL(Dummy)=0) and relationship (REL(Dummy)=1) loans. The first two columns report the mean (medians in parentheses) values for various price and non-price terms of loan contract. Panel B provides similar details for borrower-specific characteristics. The last column provides t-statistic for difference in means (z-statistic for Wilcoxon Rank sum test). All values are winsorized at the 1% and 99% level. Panel A: Loan Characteristics No Relationship REL(Dummy) = 0 (A) AISD in basis points 238.97 (250.00) AISU in basis points 37.36 (37.50) Upfront fee in basis points 76.75 (50.00) Annual fee in basis points 23.52 (12.50) Facility Size ($ millions) 125.00 (40.00) Maturity (months) 44.84 (36.00) Collateral 0.84 (1.00) Syndicate 0.78 (1.00) Panel B: Borrower Characteristics Total Assets ($ millions) 2125.67 (279.82) Coverage=(EBITDA/Interest) 36.64 (4.38) Leverage=LT Debt/Total Assets 0.25 (0.23) Profitability=EBITDA/Sales 0.14 (0.11) Tangibility=NPPE/Total Assets 0.35 (0.29) Current Ratio=CA/CL 1.96 (1.64) Market to Book 1.68 (1.31) Investment Grade 0.42 0.00 Not Rated 0.75 1.00 53 Relationship REL(Dummy) = 1 (B) 187.47 (175.00) 30.22 (25.00) 53.16 (30.00) 17.07 (12.50) 271.00 (100.00) 42.35 (36.00) 0.77 (1.00) 0.90 (1.00) t- statistic (A)-(B) (z- statistic for Wilcoxon Sum Test) 23.86*** (23.08***) 14.56*** (17.43***) 12.15*** (14.58***) 8.16*** (6.12***) -18.56*** (-34.89***) 5.84*** (5.02***) 10.43*** (10.38***) -22.95*** (-22.68***) 4074.99 (714.03) 33.50 (5.06) 0.27 (0.26) 0.16 (0.13) 0.35 (0.30) 1.90 (1.58) 1.73 (1.39) 0.51 1.00 0.62 1.00 -10.00*** (-26.94***) 0.27 (-10.04***) -7.24*** (-8.59***) -9.69*** (-13.08***) -2.30** (-10.80***) 2.50** (17.92***) -2.15** (-8.55***) -6.22*** (-6.20***) 19.86*** (19.68***) TABLE 4 Effect of lending relationships on cost of borrowing This table provides the OLS estimates (corrected for heteroscedasticity and clustering) of the following equation. AISD = β0 + βi (REL(M )) + X βi (Loan Charactersticsi ) + X βj (Borrower Charactersticsj ) + X βk (Controlk ). The dependant variable AISD is the the coupon spread over LIBOR on the drawn amount plus the annual fee in basis points. REL(M) is the measure of relationship strength, estimated in 3 different ways: REL(Dummy)(1 if there is a relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise), REL(Number)(ratio of number of deals with the lead bank(s) to total number of loans borrowed by the firm in the last 5 years before the current loan), and REL(Amount) (ratio of dollar value of deals with the lead bank(s) to total dollar value of loans borrowed by the firm in the last 5 years before the current loan). For a facility with multiple lead banks, the maximum REL(M) value among all the lead banks is used. Maturity is length in months between facility activation date and maturity date. Loan Amount is the loan facility size in millions of real year 2000 dollars. Collateral is a dummy variable that equals one if the facility has the stated attribute and zero otherwise. Log(Assets) is the natural log of book value of assets in real year 2000 dollars of the borrower as reported in the COMPUSTAT. Leverage is the ratio of book value of total debt to book value of assets. Coverage is the ratio of EBITDA to interest expenses. Profitability is the ratio of EBITDA to Sales. Tangibility is the ratio of NPPE to Total Assets. Current Ratio is the ratio of Current Assets to Current Liabilities. Market to book is the ratio of (Book value of assets-book value of equity+market value of equity) divided by book value of assets. In addition to the variables reported, the regression also includes industry dummies based on the one-digit SIC code of the borrower, dummies for stated purpose of the facility, dummies for loan type, calendar year dummies and dummies for the credit rating of the borrower, with not rated firms considered a separate group. Numbers in the parentheses are standard errors corrected for heteroscedasticity and firm level clustering. Clustering at deal level produces similar results. (*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) Const. (1) 903.21∗∗∗ (22.76) Rel(Dummy) -10.55∗∗∗ (2.44) Rel(Number) (2) 904.42∗∗∗ (22.95) (3) 903.82∗∗∗ (22.89) (4) 331.57∗∗∗ (19.40) -11.15∗∗∗ (2.47) -9.94∗∗∗ (2.72) Rel(Amount) -10.15∗∗∗ (2.68) Log(Maturity) -18.07∗∗∗ (2.54) -17.91∗∗∗ (2.54) -17.95∗∗∗ (2.54) -18.58∗∗∗ (2.46) Collateral 60.01∗∗∗ (2.87) 59.99∗∗∗ (2.87) 59.99∗∗∗ (2.87) 58.82∗∗∗ (2.70) Default Spread 34.47∗∗∗ (1.90) Log(Loan Size) -16.17∗∗∗ (1.56) -16.17∗∗∗ (1.55) -16.05∗∗∗ (1.55) -16.07∗∗∗ (1.23) Log(Assets) -9.71∗∗∗ (1.57) -10.06∗∗∗ (1.56) -10.05∗∗∗ (1.56) -9.55∗∗∗ (1.16) Log(1+coverage) -23.91∗∗∗ (1.69) -23.93∗∗∗ (1.70) -23.91∗∗∗ (1.70) -24.19∗∗∗ (1.56) Leverage 21.63∗∗ (8.73) 21.21∗∗ (8.71) 21.44∗∗ (8.72) 23.92∗∗∗ (7.98) Profitability -44.03∗∗∗ (13.65) -43.68∗∗∗ (13.64) -44.00∗∗∗ (13.62) -41.86∗∗∗ (11.29) Tangibility -12.00∗ (6.73) -11.78∗ (6.76) -11.81∗ (6.76) -13.31∗∗ (5.29) Current Ratio -2.71∗∗∗ (.74) -2.71∗∗∗ (.74) -2.71∗∗∗ (.74) -2.45∗∗∗ (.71) Market to Book -1.13 (.96) -1.12 (.96) -1.12 (.96) -1.95∗∗ (.92) Obs. R2 13158 .57 13158 .57 13158 .57 13158 .56 54 TABLE 5 Endogeneity of Lending Relationships Panel A of this table reports the average difference in AISD of: (a) relationship loans and non-relationship loans; (b) relationship loans and loans by borrowers that had lending relationships in the past but obtained the loan from a non-relationship lender; and, (c) relationship loans and loans by borrowers that had no prior lending relationships. AISD is the coupon spread on the drawn amount plus the annual fee in basis points. To examine mean AISD spread differences, we control for various borrower and lender characteristics: We compute propensity scores using the following probit model: REL = β0 + X βi (Loan Charactersticsi ) + X βj (Borrower Charactersticsj) + X βk (Controlk ). The dependent variable is REL, a dummy variable that equals one if there is a past relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise. The Loan Characteristics include log of loan size, and dummy variables for the type and purpose of the loan. Borrower characteristics include log of assets, profitability, tangibility, leverage, interest coverage, current ratio, market to book ratio, and borrower rating. Other controls include one-digit SIC code of the borrower, term spread, and default spreads prevailing at the time of the loan origination. We use the following four estimators. The NEAREST NEIGHBOR estimator chooses for each relationship loan, the n non-relationship loans with closest propensity scores and uses the arithmetic average of AISD for these n non-relationship loans. We use n = 10 and n = 50. The GAUSSIAN and EPANECHNIKOV estimators use a weighted average of AISD of non-relationship loans, with more weight given to non-relationship loans with propensity scores that are closer to the relationship loan’s propensity score. The GAUSSIAN estimator uses all the matched non-relationship loans, while for the EPANECHNIKOV estimator, we specify a propensity score bandwidth (h) that limits the sample of non-relationship loans. We specify that h = 0.01. In column (1), we compute mean difference in AISD between relationship loans and non-relationship loans by using the estimators to match relationship loans to non-relationship loans. In column (2), we compute mean difference in AISD between relationship loans and only those non-relationship loans where the borrower had relationship but chooses to obtain the loan from a non-relationship lender. In column (3), we compute mean difference in AISD between relationship loans and only those non-relationship loans where the borrowers did not have a past relationship. For all estimations, we present the sample averages. We report standard errors in parentheses that are computed by bootstrapping with 50 replications. Panel B provides the results of an instrumental variables (IV) estimation with log(1+distance) as an instrument for relationship. Distance is the spherical distance between the lead lender and the borrower. Panel C reports the results of a treatment effects model. This table provides the results of treatment effects model of the following model AISDi = Reli∗ = β0 + β1 Reli + X βk (Control)k + ǫi X γ0 + γi log(1 + distance)i + βj (Control)j + ui where the observed decision is Reli = 1 if Reli∗ > 0 and 0 otherwise, where ǫ and u are assumed to have a bivariate distribution with correlation ρ. The first column reports the estimation of Relationship. The second column reports the estimation of effect of relationship on AISD. In addition to the variables reported, the regression also includes controls for the one-digit SIC code of the borrower, stated purpose of the facility, loan type, time, and credit rating of the borrower. ***, **, * indicates significantly different than zero at the 1%, 5%, and 10% level, respectively. Panel A: Propensity Score Matching Estimation Estimator 1 2 3 Nearest Neighbor (n=10) -12.27*** (2.93)*** -13.65*** (2.64) -7.69 (4.77) Nearest Neighbor (n=50) -10.94*** (2.44) -12.52*** (2.41) -7.02* (4.25) Gaussian -11.16*** (2.37) -12.56*** (1.84) -8.64** (3.85) Epanechnikov -10.95*** (2.49) -12.53*** (1.83) -7.82** (3.55) 55 TABLE 5 Contd Panel B: Instrumental Variables Regression Estimation of effect of Relationship on AISD First Stage Regression Instrumental Variables Estimation OLS IV Constant -4.552∗∗∗ Constant 331.57∗∗∗ 422.16∗∗∗ (0.001) (19.40) (23.80) Log(1+distance) -0.058∗∗∗ Rel(Dummy) -11.15∗∗∗ -56.90∗∗ (0.011) (2.47) (25.14) Log(Assets) Log(Loan Size) Profitability Tangibility 56 Leverage Log(1+coverage) Current Ratio Market to Book Observations Pseudo R2 -0.002 (0.015) 0.175∗∗∗ (0.015) -0.108 (0.168) -0.083 (0.082) 0.443∗∗∗ (0.098) 0.096∗∗∗ (0.016) 0.010 (0.010) -0.007 (0.011) 16074 0.086 Log(Maturity) Collateral Default Spread Log(Loan Size) Log(Assets) Log(1+coverage) Leverage Profitability Tangibility Current Ratio Market to Book Obs. R2 First Stage F-statistic -18.58∗∗∗ (2.46) 58.82∗∗∗ (2.70) 34.47∗∗∗ (1.90) -16.07∗∗∗ (1.23) -9.55∗∗∗ (1.16) -24.19∗∗∗ (1.56) 23.92∗∗∗ (7.98) -41.86∗∗∗ (11.296) -13.31∗∗ (5.29) -2.45∗∗∗ (0.71) -1.95∗∗ (0.92) 13158 0.56 -15.40∗∗∗ (2.32) 71.41∗∗∗ (3.25) 35.71∗∗∗ (2.50) -11.63∗∗∗ (2.01) -12.59∗∗∗ (1.49) -24.12∗∗∗ (2.12) 55.56∗∗∗ (11.07) -57.85∗∗∗ (15.77) -20.14∗∗∗ (7.27) -1.70∗∗ (0.80) -3.23∗∗∗ (1.02) 13158 0.49 128.18∗∗∗ TABLE 5 Contd Panel C: Treatment Effects Model Relationship AISD Constant -0.22 397.62∗∗∗ (0.33) (18.05) -16.91∗∗∗ Rel(Dummy) (4.07) Log(1+distance) -0.055∗∗∗ (0.01) Default Spread Log(Loan Size) Log(Assets) Log(1+coverage) 0.05∗ 35.86∗∗∗ (0.03) (2.46) 0.14∗∗∗ -15.01∗∗∗ (0.02) (1.47) 0.04∗∗ -12.19∗∗∗ (0.02) (1.45) 0.11 ∗∗∗ (0.02) Leverage Profitability Tangibility Current Ratio Market to Book -25.81∗∗∗ (1.86) 0.559∗∗∗ 45.86∗∗∗ (0.11) (9.67) -0.16 -57.34∗∗∗ (0.19) (15.53) -0.15∗ -17.13∗∗ (0.09) (7.17) 0.01 -1.81∗∗ (0.01) (0.79) 0.003 -3.13∗∗∗ (0.01) (1.03) -12.79∗∗∗ Log(Maturity) (1.73) 72.07∗∗∗ Collateral (3.04) Obs. 13158 ρ 0.042 LR test of ρ = 0 χ2 (1)=4.84 Probability > χ2 =0.028 57 13158 TABLE 6 Borrower Information Opacity and Benefits of Relationship This table provides the OLS regression (corrected for heteroscedasticity and clustering) estimates of the following equation. AISD = β0 + β1 (REL(M )) + β2 (Borrower Inf ormation Opacity) + β3 (REL(M )) × (Borrower Inf ormation Opacity) X X X βj (Borrower Charactersticsj ) + βk (Controlk ). βi (Loan Charactersticsi ) + + The dependant variable AISD is the the coupon spread on the drawn amount plus the annual fee in basis points. REL(Dummy) equals 1 if there is a relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise. Maturity is length in months between facility activation date and maturity date. The Loan Amount is the loan facility size in millions of real year 2000 dollars. Collateral, is a dummy variable that equals one if the facility has the stated attribute and zero otherwise. Log(Assets) is the natural log of book value of assets of the borrower in millions of year 2000 real dollars as reported in the COMPUSTAT. Leverage is the ratio of book value of total debt to book value of assets. Coverage is the ratio of EBITDA to interest expenses. Profitability is the ratio of EBITDA to Sales. Tangibility is the ratio of NPPE to Total Assets. Current Ratio is the ratio of Current Assets to Current Liabilities. Market to book is the ratio of (Book value of assets-book value of equity+market value of equity) divided by book value of assets. BORROWER INFORMATION OPACITY is measured by six different proxies: Borrower’s Size (Log of Assets), if the borrower is not rated (Not Rated Dummy), if the borrower is in the S&P 500 index, number of analysts following the borrower firm, a microstructure measure of information asymmetry ASY, developed by Bharath, Pasquariello, and Wu (2007), discretionary accruals computed according to Dechow and Dichev (2002). Big three is a dummy variable if the lead bank is either Citibank, JP Morgan Chase, or Bank of America. In addition to the variables reported, the regression also includes industry dummies based on the one-digit SIC code of the borrower, dummies for stated purpose of the facility, dummies for loan type, calendar year dummies, and dummies for the credit rating of the borrower, with not rated firms considered a separate group. Numbers in the parentheses are standard errors corrected for heteroscedasticity and firm-level clustering. Clustering at deal level produces similar results. (*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) (Continued on next page) 58 Table 6 (continued) Const. (1) 927.16∗∗∗ (23.66) (2) 902.33∗∗∗ (22.75) (3) 436.46∗∗∗ (19.88) (4) 903.75∗∗∗ (23.22) (5) 909.74∗∗∗ (24.54) (6) 518.53∗∗∗ (18.49) (7) 903.09∗∗∗ (22.67) Rel(Dummy) -43.77∗∗∗ (8.94) -2.35 (3.65) -14.42∗∗∗ (2.72) -14.81∗∗∗ (3.07) -9.99∗∗∗ (2.71) -6.39∗ (3.50) -14.39∗∗∗ (2.74) Rel(Dummy) × Log(Assets) 5.50∗∗∗ (1.37) Rel(Dummy) × Not Rated -12.37∗∗∗ (4.73) Rel(Dummy) × SP500 26.94∗∗∗ (6.27) Rel(Dummy) × No. Analysts .93∗∗∗ (.30) Rel(Dummy) × ASY -3.23∗∗ (1.65) Rel(Dummy) × Accruals -62.60∗ (34.77) Rel(Dummy) × Big-3 bank 22.98∗∗∗ (5.28) SP500 -12.26 (8.17) No. of Analysts -.89∗∗∗ (.32) ASY 2.14∗ (1.29) Discretionary Accruals 99.90∗∗∗ (33.57) Big-3 bank -20.59∗∗∗ (4.64) Log(Maturity) -18.27∗∗∗ (2.55) -18.06∗∗∗ (2.54) -18.07∗∗∗ (2.63) -18.16∗∗∗ (2.55) -18.49∗∗∗ (2.98) -20.05∗∗∗ (3.11) -18.04∗∗∗ (2.55) Collateral 59.78∗∗∗ (2.86) 59.80∗∗∗ (2.86) 57.17∗∗∗ (2.90) 59.82∗∗∗ (2.86) 55.92∗∗∗ (3.14) 57.64∗∗∗ (3.46) 59.80∗∗∗ (2.86) Log(Assets) -13.53∗∗∗ (1.79) -9.73∗∗∗ (1.56) -10.35∗∗∗ (1.59) -9.31∗∗∗ (1.62) -8.72∗∗∗ (1.66) -7.88∗∗∗ (1.86) -9.81∗∗∗ (1.57) Log(Loan Size) -16.15∗∗∗ (1.55) -16.15∗∗∗ (1.55) -16.71∗∗∗ (1.60) -16.10∗∗∗ (1.55) -17.70∗∗∗ (1.63) -17.89∗∗∗ (1.82) -15.89∗∗∗ (1.55) Log(1+coverage) -23.52∗∗∗ (1.70) -23.76∗∗∗ (1.69) -24.48∗∗∗ (1.73) -23.70∗∗∗ (1.70) -25.40∗∗∗ (1.83) -29.09∗∗∗ (2.37) -23.84∗∗∗ (1.69) Leverage 22.77∗∗∗ (8.74) 22.04∗∗ (8.74) 18.69∗∗ (8.86) 21.27∗∗ (8.74) 12.75 (9.25) 21.19∗∗ (10.39) 21.13∗∗ (8.70) Profitability -43.81∗∗∗ (13.61) -44.26∗∗∗ (13.64) -43.08∗∗∗ (14.40) -43.49∗∗∗ (13.70) -41.30∗∗∗ (15.62) -16.89 (16.71) -43.52∗∗∗ (13.61) Tangibility -12.10∗ (6.69) -11.95∗ (6.73) -2.35 (6.91) -11.93∗ (6.72) -7.10 (7.11) -6.32 (7.70) -11.94∗ (6.73) Current Ratio -2.65∗∗∗ (.73) -2.72∗∗∗ (.73) -2.73∗∗∗ (.77) -2.68∗∗∗ (.73) -3.13∗∗∗ (.82) -3.09∗∗∗ (1.04) -2.71∗∗∗ (.73) Market to Book -1.23 (.95) -1.15 (.96) -1.49 (.98) -1.07 (.97) -1.19 (1.00) -1.06 (1.36) -1.11 (.96) Obs. R2 13158 0.57 13158 0.57 12180 0.58 13158 0.57 10744 0.59 9170 0.60 13158 0.57 59 TABLE 7 Role of Relationships as a Bank’s Commitment to Monitor This table provides the OLS regression (corrected for heteroscedasticity and clustering) estimates of the following equation. AISD = β0 + β1 (REL(M )) + β2 (Bank M onitoring Incentive) + β3 (REL(M )) × (Bank M onitoring Incentive) X X X βi (Loan Charactersticsi ) + βj (Borrower Charactersticsj ) + βk (Controlk ). + The dependant variable AISD is the the coupon spread on the drawn amount plus the annual fee in basis points. REL(Dummy) equals 1 if there is a relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise. Maturity is length in months between facility activation date and maturity date. The Loan Amount is the loan facility size in millions of real year 2000 dollars. Collateral, is a dummy variable that equals one if the facility has the stated attribute and zero otherwise. Not Rated equals one if the borrower does not have rating from S&P and zero otherwise. Log(Assets) is the natural log of book value of assets of the borrower in millions of year 2000 real dollars as reported in the COMPUSTAT. Leverage is the ratio of book value of total debt to book value of assets. Coverage is the ratio of EBITDA to interest expenses. Profitability is the ratio of EBITDA to Sales. Tangibility is the ratio of NPPE to Total Assets. Current Ratio is the ratio of Current Assets to Current Liabilities. Market to book is the ratio of (Book value of assets-book value of equity+market value of equity) divided by book value of assets. Syndicate moral hazard, i.e., the incentive to monitor the borrower is measured by 3 different proxies: number of members in the syndicate, lead bank allocation of the loan in percent, the loan concentration in the syndicate measured by the Herfindahl-Hirschmann index following Sufi (2007). Covenant Index assumes the value between zero and five with the presence of each of five different covenants coded as a one and zero otherwise and summed up (Bradley and Roberts, 2004). In addition to the variables reported, the regression also includes industry dummies based on the one-digit SIC code of the borrower, dummies for stated purpose of the facility, dummies for loan type, calendar year dummies, and dummies for the credit rating of the borrower, with not rated firms considered a separate group. Numbers in the parentheses are standard errors corrected for heteroscedasticity and firm-level clustering. Clustering at deal level produces similar results. (*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) (Continued on next page) 60 Table 7 Panel A Const. (1) 569.22∗∗∗ (2) 642.95∗∗∗ (3) 898.58∗∗∗ (4) 855.71∗∗∗ (5) 834.47∗∗∗ (60.96) (119.35) (23.18) (38.02) (43.21) Rel(Dummy) -18.19∗∗∗ -14.44∗∗ -20.27∗∗∗ 6.25 3.57 (6.12) (7.36) (3.70) (4.96) (4.52) 8.59∗∗∗ Rel(Dummy) × Log(Syndicate Size) (2.02) -7.63∗∗∗ Log(Syndicate Size) (2.17) -.26∗∗∗ Rel(Dummy) × Lead Bank Allocation (.09) .49∗∗∗ Lead bank Allocation (.09) -.002∗∗∗ Rel(Dummy) × Loan HHI (.0008) .005∗∗∗ Loan HHI (.001) Log(Maturity) Collateral Log(Assets) Log(Loan Size) Log(1+coverage) -16.21∗∗∗ -19.15∗∗∗ -17.82∗∗∗ -15.39∗∗∗ -14.67∗∗∗ (5.81) (6.44) (2.57) (3.61) (3.75) ∗∗∗ ∗∗∗ 63.45 64.70 59.86 60.82 60.11∗∗∗ (7.75) (8.53) (2.88) (3.62) (3.72) Profitability Tangibility Current Ratio Market to Book Obs. R2 ∗∗∗ -21.10∗∗∗ -17.00∗∗∗ -9.82∗∗∗ -15.50∗∗∗ -15.33∗∗∗ (3.75) (4.90) (1.59) (1.80) (1.89) -20.29∗∗∗ -25.61∗∗∗ -15.34∗∗∗ -11.37∗∗∗ -10.49∗∗∗ (4.24) (5.04) (1.63) (2.00) (2.21) -15.07 ∗∗∗ -13.89 (2.71) Leverage ∗∗∗ ∗∗∗ (2.94) -23.58 ∗∗∗ (1.70) -24.18 ∗∗∗ (2.09) -24.24∗∗∗ (2.15) 54.82∗∗ 43.82 22.12∗∗ 17.91 23.59∗∗ (22.46) (26.75) (8.73) (11.40) (11.89) -110.31∗∗∗ -122.42∗∗∗ -43.73∗∗∗ -30.44∗ -35.52∗∗ (42.47) (43.45) (13.58) (16.50) (16.94) -30.57 -18.41 -11.93∗ -13.92 -14.13 (21.26) (23.54) (6.73) (8.76) (9.08) -4.72∗∗ -5.49∗∗ -2.74∗∗∗ -3.80∗∗∗ -4.16∗∗∗ (2.12) (2.23) (.73) (1.22) (1.24) .12 .52 -1.14 -1.14 -.90 (2.25) (2.36) (.95) (1.17) (1.19) 1669 0.43 1184 0.44 13158 0.57 6064 0.58 5731 0.58 61 Table 7 Panel B Const. Rel(Dummy) Rel(Dummy) * Collateral (1) 902.95∗∗∗ (2) 911.75∗∗∗ (3) 909.15∗∗∗ (4) 909.43∗∗∗ (22.79) (23.14) (23.18) (23.15) -1.83 -9.74∗∗∗ -10.92∗∗∗ -8.44∗∗∗ (3.45) (3.11) (2.44) (2.88) -3.00∗∗∗ -2.22∗∗ (.81) (.95) -15.28∗∗∗ (4.73) Rel(Dummy) * Covenant Index -.85 (1.68) Market to Book * Covenant Index Rel(Dummy) * Market to Book * Covenant Index -1.16 (.75) ∗∗∗ 7.18 11.22 11.29∗∗∗ (1.87) (1.90) (1.91) -18.21∗∗∗ -18.91∗∗∗ -18.63∗∗∗ -18.64∗∗∗ (2.55) (2.53) (2.52) (2.52) 70.72∗∗∗ 49.12∗∗∗ 49.13∗∗∗ 49.16∗∗∗ (4.72) (3.65) (3.65) (3.64) Covenant Index Log(Maturity) Collateral Log(Assets) Log(Loan Size) Log(1+coverage) Leverage ∗∗∗ -9.58 -9.45 (1.56) (1.58) Tangibility Current Ratio Market to Book Obs. R2 -9.81 ∗∗∗ (1.57) -9.77∗∗∗ (1.57) -16.22∗∗∗ -16.69∗∗∗ -16.44∗∗∗ -16.47∗∗∗ (1.55) (1.57) (1.57) (1.57) -23.79∗∗∗ -24.30∗∗∗ -24.18∗∗∗ -24.20∗∗∗ (1.69) (1.72) (1.73) (1.73) 22.18 ∗∗ (8.73) Profitability ∗∗∗ ∗∗∗ ∗∗ ∗∗ 17.78 18.66 18.87∗∗ (8.76) (8.74) (8.76) -43.74∗∗∗ -44.63∗∗∗ -44.83∗∗∗ -44.69∗∗∗ (13.64) (13.64) (13.64) (13.63) -12.07∗ -10.27 -10.54 -10.74 (6.71) (6.68) (6.68) (6.67) -2.66∗∗∗ -2.52∗∗∗ -2.58∗∗∗ -2.53∗∗∗ (.73) (.73) (.75) (.75) -1.11 -.98 1.59 1.52 (.95) (.98) (1.51) (1.49) 13158 .57 13158 .57 13158 .58 13158 .58 62 Table 8 Lending Relationships and Probability of Pledging Collateral This table provides the logit regression estimates (corrected for heteroscedasticity and clustering) of the following equation. COLLAT ERAL = β0 + β1 (REL(M )) + β2 Log(Loan Amount) + β3 (Leverage) + β4 (T angibility) + β5 (M arket to Book) + β6 (Loan Concentration) X β7 (Log(M aturity) + β8 (N ot Rated) + βk (Controlk ). + The dependant variable COLLATERAL is a dummy variable that equals 1 if a loan facility is secured by collateral and 0 otherwise. REL(M) is the measure of relationship strength, estimated in 3 different ways: REL(Dummy)(1 if there is a relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise), REL(Number)(ratio of number of deals with the lead bank(s) to total number of loans borrowed by the firm in the last 5 years before the current loan), and REL(Amount) (ratio of dollar value of deals with the lead bank(s) to total dollar value of loans borrowed by the firm in the last 5 years before the current loan). For a facility with multiple lead banks, the maximum REL(M) value among all the lead banks is used. The Loan Amount is the loan facility size in millions of real year 2000 dollars. Leverage is the ratio of book value of total debt to book value of assets. Tangibility is the ratio of NPPE to Total Assets. Market to book is the ratio of (Book value of assets-book value of equity+market value of equity) divided by book value of assets. Loan Concentration is the ratio of that loan facility amount to sum existing debt and the amount of loan facility. In addition to the variables reported, the regression also includes industry dummies based on the one-digit SIC code of the borrower, dummies for stated purpose of the facility, calendar year dummies, and dummies for the credit rating of the borrower, with not rated firms considered a separate group. Numbers in the parentheses are standard errors corrected for heteroscedasticity and firm-level clustering. Clustering at deal level produces similar results.(*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) Const. (1) -18.15∗∗∗ (.62) Rel(Dummy) -.22∗∗∗ (.05) (2) -18.15∗∗∗ (.63) Rel(Number) (3) -18.16∗∗∗ (.63) -.26∗∗∗ (.06) Rel(Amount) -.25∗∗∗ (.06) Log(Loan Size) -.47∗∗∗ (.03) -.47∗∗∗ (.03) -.47∗∗∗ (.03) Leverage 2.93∗∗∗ (.27) 2.96∗∗∗ (.27) 2.96∗∗∗ (.27) Tangibility -.48∗∗∗ (.16) -.48∗∗∗ (.16) -.48∗∗∗ (.16) Market to Book -.12∗∗∗ (.03) -.12∗∗∗ (.03) -.12∗∗∗ (.03) Loan Concentration 1.54∗∗∗ (.16) 1.57∗∗∗ (.16) 1.57∗∗∗ (.16) Log(Maturity) .25∗∗∗ (.03) .26∗∗∗ (.03) .25∗∗∗ (.03) Obs. Pseudo R2 15510 0.26 15510 0.26 15510 0.26 63 TABLE 9 Lending Relationships and Loan Maturity This table provides the OLS regression (corrected for heteroscedasticity and clustering) estimates of the following equation. Log(M aturity) = + β0 + β1 (REL(M )) + β2 (Log(Loan Amount)) + β3 (Leverage) + β4 (Log(Assets)) + β5 (M arket to Book) X β6 (Log(Asset M aturity)) + β7 (Regulated) + β8 (Collateral) + βk (Controlk ). The dependant variable Log(Maturity) is the natural log of the stated maturity of the loan facility (measured as length in months between facility activation date and maturity date). REL(M) is the measure of relationship strength, estimated in 3 different ways: REL(Dummy)(1 if there is a relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise), REL(Number)(ratio of number of deals with the lead bank(s) to total number of loans borrowed by the firm in the last 5 years before the current loan), and REL(Amount) (ratio of dollar value of deals with the lead bank(s) to total dollar value of loans borrowed by the firm in the last 5 years before the current loan). For a facility with multiple lead banks, the maximum REL(M) value among all the lead banks is used. The Log(Assets) is the natural log of the book value of assets of the borrower. Loan Amount is the loan facility size. Leverage is the ratio of book value of total debt to book value of assets. Market to book is the ratio of (Book value of assets-book value of equity+market value of equity) divided by book value of assets. Asset Maturity is the weighted average of current assets divided by cost of goods sold, and Net PPE divided by depreciation and amortization - as defined in Barclay, Marx, and Smith (2003). Regulated Industry is a dummy variable that equals one for firms in the Utilities industry under the Fama-French industry classification and zero otherwise. Term Spread is the difference in yields between 1-year and 10-year US Government Bonds at the time of loan activation. In addition to the variables reported, the regression also includes industry dummies based on the one-digit SIC code of the borrower, dummies for stated purpose of the facility and calendar year dummies. Numbers in the parentheses are standard errors corrected for heteroscedasticity and firm-level clustering. Clustering at deal level produces similar results. (*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) Const. (1) 2.58∗∗∗ (.10) Rel(Dummy) -.08∗∗∗ (.02) (2) 2.65∗∗∗ (.10) Rel(Dummy) × High -.10∗∗∗ (.02) Rel(Dummy)× Middle -.01 (.02) Rel(Dummy) × Low -.15∗∗∗ (.03) (3) 3.42∗∗∗ (.23) (4) 3.73∗∗∗ (.23) (5) 2.57∗∗∗ (.13) (6) 2.54∗∗∗ (.10) -.09∗∗ (.04) -.05∗ (.03) -.09∗∗∗ (.02) -.08∗∗∗ (.02) Log(Loan Size) .14∗∗∗ (.009) .13∗∗∗ (.009) .09∗∗∗ (.02) .13∗∗∗ (.02) .14∗∗∗ (.01) .14∗∗∗ (.009) Leverage .41∗∗∗ (.04) .38∗∗∗ (.04) .36∗∗∗ (.11) .38∗∗∗ (.08) .39∗∗∗ (.05) .41∗∗∗ (.04) Log(Assets) -.05∗∗∗ (.009) -.05∗∗∗ (.009) -.03∗ (.02) -.08∗∗∗ (.02) -.03∗∗∗ (.01) -.05∗∗∗ (.009) Market to Book -.005 (.005) -.005 (.005) .006 (.01) .02 (.02) -.008 (.006) -.005 (.005) Log (Asset Maturity) .03∗∗∗ (.01) .04∗∗∗ (.01) .03 (.02) -.04 (.02) .06∗∗∗ (.01) .03∗∗∗ (.01) Regulated -.47∗∗∗ (.06) -.47∗∗∗ (.05) -.30∗∗∗ (.06) -.52∗∗∗ (.13) -.49∗∗∗ (.10) -.47∗∗∗ (.06) Collateral .12∗∗∗ (.02) .12∗∗∗ (.02) .26∗∗∗ (.06) .14∗∗∗ (.03) .08∗∗∗ (.02) .12∗∗∗ (.02) Term Spread Obs. R2 .0004∗∗ (.0002) 15636 0.25 15636 0.25 3174 0.33 64 3289 0.27 9173 0.21 15636 0.25 TABLE 10 Instrumental Variables Estimation of Loan Spread, Maturity and Collateral This table provides estimations of following system of equations using instrumental variables to estimate the impact of past relationships on Loan Spread, Maturity and Collateral. AISD = γA (REL) + γAC (Collateral) + γAM (Log(M aturity)) + XA βA + ǫA Collateral = γC (REL) + γCM (Log(M aturity)) + XC βC + ǫC Log(M aturity) = γM (REL) + γM C (Collateral) + XM βM + ǫM AISD is the the coupon spread over LIBOR on the drawn amount plus the annual fee. Log(Maturity) is the natural log of the stated maturity of the loan facility (Measured as length in months between facility activation date and maturity date). Collateral is a dummy variable that equals 1 if the loan was secured and zero otherwise. REL(Dummy)(1 if there is a relationship with any of the lead banks in the last 5 years before the present loan and 0 otherwise). Default spread is measured as the difference between the yields on Moody’s seasoned corporate bonds with Baa rating and 10-year U.S. government bond. Asset Maturity is the weighted average of current assets divided by cost of goods sold, and Net PPE divided by depreciation and amortization - as defined in Barclay, Marx, and Smith (2003). Regulated Industry is a dummy variable that equals one for firms in the Utilities industry under the Fama-French industry classification and zero otherwise. Loan Concentration is the ratio of that loan facility amount to sum existing debt and the amount of loan facility. Durbin Wu-Hausmann Chi-sq test is the test of null hypothesis that maturity and collateral are indeed exogenous. Rejection of null implies they are endogenous. Wald test provides similar test for IV probit model for collateral. To test for instrument relevance, for AISD regression since we have 2 endogenous regressors we estimate the Cragg-Donald statistic and compare it against critical values in table 1 of Stock and Yogo (2005). For Log(Maturity) regression we report the the first stage F-statistic as well as partial R2 since it has only one endogenous variable. For both AISD and Log(maturity) IV regressions, we report Anderson LR test of the null hypothesis that our instruments and endogenous variables are not correlated. We reject the null (p-value =0.00) implying our instruments are correlated with the endogenous variables. Since we have two endogenous variables but four instruments for the spread equation, we are able to estimate the Hansen-J statistic for over-identification restrictions. These are joint tests of the null hypothesis that the correct model is specified and the orthogonality conditions are met (correlations between the instruments and the error term is zero). Our test statistic does not reject the null for AISD and Log(Maturity) regressions implying exogeneity of the instruments. For both AISD and Log(Maturity) IV regressions, numbers in the parentheses are standard errors corrected for heteroscedasticity and firm-level clustering. For IV probit cluster-corrected standard errors cannot be computed so we use boot strapping with 50 replications to estimate the standard errors. (*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) (Continued on next page) 65 Dependent Variable REL(Dummy) AISD OLS -10.55*** (2.44) IV -14.35** (5.73) IV -12.79** (5.73) REL(Dummy)*High REL(Dummy)*Low Collateral Log(Maturity) IV -0.10*** (0.02) -0.01 (0.02) -0.15*** (0.03) REL(Dummy)*Middle Log(Maturity) OLS -18.07*** (2.54) 60.01*** (2.87) Default Spread -151.6*** (42.70) 406.2*** (68.09) 15.44*** (7.62) Average Spread -146.13*** (42.04) 408.2*** (64.94) Regulated 0.15*** (0.02) 0.94*** (0.15) 0.90*** (0.09) 1.00*** (0.07) As in Column 1 of Table 8 -0.02 (0.03) 0.07*** (0.03) -0.10*** (0.04) 0.12*** (0.02) 1.67*** (0.97) 0.04*** (0.01) -0.47*** (0.05) 0.65*** (0.15) -0.24*** (0.77) Loan Concentartion Tests of edogeneity: Durbin-Wu-Hausman chi-sq test p-value Wald Chi-Square test staistic p-value Weak identification statistics:: Cragg-Donald F-Stat Stock and Yogo (2004) critical value First Stage F statistic Partial R-Sqaure Anderson -LR statistic p-value Instrument exogeneity test: Hansen-J statistic p-value Collateral IV Probit -0.07*** (0.31) 0.68*** (0.11) Log(Asset Maturity) Other Controls Probit -0.13*** (0.03) As in Column 1 of Table 4 As in Column 1 of Table 4 As in Column 2 of Table 9 469.25 0.00 546.46 0.00 1170 0.00 29.56 0.00 14.57 15.05 11.03 11.03 58.24 0.00 60.17 0.00 323.27 8.16% 1289 0.00 4.132 0.13 3.967 0.14 0.001 0.98 66 TABLE 11 Lending Relationships and Access to Loans The first three specifications of this table provides the OLS estimates (corrected for heteroscedasticity) of the following equation. Loan Amount Assets = β0 + β1 (REL(M )) + X βi (Loan Charactersticsi ) + X βj (Borrower Charactersticsj ) + X βk (Controlk ). Loan Amount The next three specifications are essentially the same but the dependent variable is Long . REL(M) is the measure of T erm Debt relationship strength, estimated in 3 different ways (we report it only for REL(M) constructed over a 5-year look back window. We report results for 3 different relationship measures. The Loan Amount is the loan facility size in millions of real year 2000 dollars. Profitability is the ratio of EBITDA to Sales. Tangibility is the ratio of PPE to Total Assets. Current Ratio is the ratio of Current Assets to Current Liabilities. Market to book is the ratio of (Book value of assets-Book value of equity+market value of equity) divided by book value of assets. In addition to the variables reported, the regression also includes industry dummies based on the one-digit SIC code of the borrower, dummies for stated purpose of the facility, calendar year dummies and dummies for the credit rating of the borrower, with not rated firms considered a separate group. Numbers in the parentheses are standard errors corrected for heteroscedasticity and firm-level clustering. Clustering at deal level produces similar results.(*** Significant at one percent level, ** Significant at five percent level ,* Significant at 10 percent level) Const. (1) .63∗∗∗ (.03) Rel(Dummy) .008∗∗ (.004) Rel(Number) (2) .63∗∗∗ (.03) (3) .63∗∗∗ (.03) (4) 1.43∗∗∗ (.06) (5) 1.42∗∗∗ (.06) (6) 1.42∗∗∗ (.06) .02∗∗∗ (.007) .009∗∗ (.004) .03∗∗∗ (.007) Rel(Amount) .01∗∗∗ (.004) .03∗∗∗ (.007) Log(Assets) -.06∗∗∗ (.002) -.06∗∗∗ (.002) -.06∗∗∗ (.002) -.10∗∗∗ (.003) -.10∗∗∗ (.003) -.10∗∗∗ (.003) Log(1+coverage) .001 (.002) .001 (.002) .001 (.002) .13∗∗∗ (.005) .13∗∗∗ (.005) .13∗∗∗ (.005) Log(Syndicate Size) .06∗∗∗ (.002) .06∗∗∗ (.002) .06∗∗∗ (.002) .07∗∗∗ (.004) .07∗∗∗ (.004) .07∗∗∗ (.004) Profitability .13∗∗∗ (.02) .13∗∗∗ (.02) .13∗∗∗ (.02) -.34∗∗∗ (.04) -.34∗∗∗ (.04) -.34∗∗∗ (.04) Tangibility -.02∗∗ (.01) -.02∗∗ (.01) -.02∗∗ (.01) -.15∗∗∗ (.02) -.15∗∗∗ (.02) -.15∗∗∗ (.02) Current Ratio .0002 (.001) .0002 (.001) .0002 (.001) -.004 (.002) -.004 (.002) -.004 (.002) Market to Book .01∗∗∗ (.003) .01∗∗∗ (.003) .01∗∗∗ (.003) .008∗∗ (.003) .008∗∗ (.003) .008∗∗ (.003) Obs. R2 15218 0.32 15218 0.32 15218 0.32 14626 0.42 14626 0.42 14626 0.42 67 Figure 1 Relationship Benefits and Borrower Size This figure displays the price break in basis points for a relationship borrower in each asset size decile compared to a similar borrower with no relationship in the same asset size decile. Asset size is measured in real year 2000 dollars. The estimate and statistical significance (t-stat) is assessed using specification 1 of Table 6. Relationship benefits measured as savings on AISD 6 10 5 4 0 -1 -4 2 -8 -10 0 -16 -20 t-stat Basis Points -11 -14 -20 -24 -2 Relationship benefits are insiginificant -30 -4 -40 -6 Relationship benefits are significant -43 -8 -50 min 10 20 30 40 50 60 Asset Size, Percentile Relationship Benefits 68 t-stat 70 80 90