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