GetWhyfirms issue callable bonds: Hedging investment uncertainty

8/10/2019 GetWhyfirms issue callable bonds: Hedging investment uncertainty

1/20


2/20

In the rst part of the paper, we develop a theory on a rm's ex ante choice between issuing a callable or non-callable bond, its

ex post decisions whether to call back a callable bond, and whether to refund it. On the one hand, our theory explains the existing

empirical ndings in the current literature, such as the lack of refunding of called bonds. On the other hand, it produces a variety of

novel testable hypotheses, which we examine empirically in the second part of the paper.

In our model, an equity-value-maximizingrm needs to raise money to invest in a current project and possibly a future project.

The current project has a positive NPV but it is uncertain whether the future project has a positive NPV. The rm decides whether

to issue a callable bond or a non-callable bond to competitive investors. After the current project generates a cash ow, the rm

and the investors observe more information about the future project. Based on the new information, the rm then decides

whether or not to invest the cash in the future project. Because the rm tries to maximize its equity value, the investment decision

may not be efcient if the bond is non-callable. More specically, the rm may want to invest in a negative NPV but risky future

project. This is because although investing in the project will lower the rm's value, it will lower the bond value even more and

equity holders can capture the difference. Anticipating that situation, investors would pay a lower price (or equivalently demand a

higher yield) for the rm's bond when it is issued than they would if the rm could commit to an efcient investment decision.

This is the well-known risk-shifting problem rst studied byJensen and Meckling (1976).

Issuing a callable bond may alleviate this risk-shifting problem. The key point is that a callable bond gives the issuing rm an

option to reduce its debt obligation if it nds out that the future project has a negative NPV. If the rm's bond is non-callable, as

discussed above, the rm may still want to invest in the project. Instead, if the rm has an option to buy back the bond at a lower

price than its value, the rm may have an incentive to not invest in the negative NPV project but pay out cash by calling back the

bond. The reason is that now the debt obligation is reduced so that the rm can keep a larger portion of its value, most of which

would go to the bond holders if it is a non-callable bond. In other words, a callable bond essentially enables the bond holders to

bribe the rm into making an efcient investment decision.

There is, however, a cost associated with issuing a callable bond. When the future project turns out to be good, the rm would

invest in the project. If the project is better than good, the rm then would want to call back the bondand refundit at a lower cost.

In this case, however, therm incurs a refunding cost.3 Therefore, therm faces the following trade-off when it decides whether to

issue a callable bond or a non-callable bond. The benet of issuing a callable bond is that it would reduce the agency cost of debt if

the investment opportunities turn out to be bad. The cost is that the rm would incur the refunding cost if the investment

opportunities turn out to be good. This implies that a rm expecting better investment opportunities would issue a non-callable

bond while it would issue a callable bond if it is expecting poorer investment opportunities.

Our model also characterizes therm's behavior after it issues a callable bond. First, if the rm nds out that itsfuture project is

bad, it would not invest in the project but call back the bond without refunding it. We thus provide an explanation to the observed

lack of refunding of called bonds discussed above. Secondly, if the rmnds out that the future project is good, it would invest in

the project, call back the bond, and refund it at a lower cost. Finally, if the rm nds out that its future project is mediocre, it would

choose to invest in the project without calling back the bond. This is because i) the project has a positive NPV so it is worth

continuing; and ii) the benet from refunding the bond is not high enough to offset the refunding cost.

Our analysis yields a variety of testable hypotheses that differentiate our theory from the alternative theories in the existing

literature. In the secondpart of thepaper, we testthosehypotheses empirically. In the ex ante (atissue)study, we examinethe relation

between a rm's decision of issuing a callable bond versus a non-callable bond and its expected future investment opportunities,

leverage ratio, and investment risk. In the ex post (at call) study, we examine the relation between a rm's current investment

performance and its decision whether or not to call back the bond, along with whether or not to refund it. We nd strong empirical

support for our theory. We nd that a rm expecting worse future investment opportunities and/or with higher leverage ratio and

investment risk is more likely to issue a callable bond. As a rm calls back its bond, the rm with the poorest performance and the

lowest investment activity is not likely to refund a call. In contrast, a rm with the best performance and the highest investment

activity is likely to refund it. A rm with mediocre performance and investment activity tends to not call their bonds. Our ndings are

also economically signicant. We estimate, forexample,thatan increase of onestandard deviation in the market/book ratio (proxyfor

future investment opportunities) corresponds to a 36% decrease in therm's probability to issue a callable bond versus a non-callable

one. In addition, we nd that, as a rm calls back a bond, a non-refunding call is associated with poorer performance and lower

investment activity. A decrease of one standard deviation in its ROA corresponds to a 15% decrease in the

rm's probability ofrefunding its called bond. Our ndings are robust to various model specications and different measures of key variables.

The rest of the paper is organized as follows.Section 2is a review of the relevant literature. In Section 3,we use a numerical

example to develop the theoretical argument that a rm can use callable bonds to reduce the risk-shifting problem. Empirical

hypotheses are also derived. A formal model is available upon request. Section 4describes our data, sample, and variables. In

Section 5, we examine the hypotheses concerning the likelihood of issuing callable versus non-callable bonds. In Section 6, we test

the hypotheses concerning the likelihood to call with refund, call without refund, and not call. Section 7concludes.

2. Literature review

The literature offers ve theories explaining why a rm issues a callable bond. The rst is the hedging interest rate risk theory in

which a callable bond provides a rm with theopportunity to refund at a lower interest rate (Pye, 1966). The second is the signaling

3 Refunding costs can be signicant for some rms.Gande et al. (1999)document an average gross spread of 2.5% for junk bonds issued from 1985 to 1996.

589Z. Chen et al. / Journal of Corporate Finance 16 (2010) 588607
http://-/?-http://-/?-


3/20

theory in which a callable bond allows a higher quality rm to reduce the cost associated withasymmetric information (Robbins and

Schatzberg, 1986, 1988). The reason is that even though a higher quality rm has to issue a bond at a lower price due to asymmetric

information, it can capture the price appreciations by calling back and refunding the bond after its true quality is revealed. The third

explanation is the resolving debt overhang theory, which indicates thata callable bond allows the issuing rm to overcome thedebt

overhang problem, as identiedby Myers (1977). If it suffers from a debt overhang problem, a rm (acting in theinterest of itsequity

holders) would not invest in positive NPV projects because part of the benets from the new projects would go to the existing bond

holders. One way to resolve this underinvestment problem is to allow the rm to call back its outstanding debt at the time of

investment and reissue debt that reects the improved prospects of the rm (Bodie and Taggart, 1978). The fourth explanation,

removing restrictive covenants theory,posits that a callable bond allows a rm to remove undesirable restrictive covenants in the

bond indentures so that the rm can engage in value-adding activities that are otherwise impossible (Vu, 1986).

The last explanation is that a rm issues a callable bond to reduce the risk-shifting problem. Equity holders can expropriate

wealth from bondholders by increasing the risk of the rm.Barnea et al. (1980)show that because the call option value of a

callable bond declines as the rm value decreases, equity holders will have less incentive to transfer wealth. We call it the

reducing risk-shifting theory.

Our theory differs from the existing theories in two important ways. First, our theory provides an explanation for why some

rms refund their called bonds but others don't. Second, our model formally studies a rm's trade-off between issuing a callable

bond versus a non-callable bond.

There is a small empirical literature on callable bonds (e.g., Vu, 1986; Kish and Livington, 1992; Crabbe and Helwege, 1994;

King and Mauer, 2000; Guntay, et al., 2004). The studies provide mixed evidence for each of the ve explanations that explain why

a rm issues a callable bond.

Overall, we think our study makes three important contributions to the current literature. First, it derives rms' equilibrium

decisions whether to issue callable bonds or non-callable bonds, when to call back the callable bonds, and whether to refund them.

Secondly, it documents empirical ndings that are consistent with our theory, but inconsistent with other theories. Lastly, to the

best of our knowledge, our study is the rst to examine a rm's commitment to payout cash by calling back its bond under poor

performance conditions.

3. Theoretical analysis

3.1. Our model

For simplicity, we use a numerical example to demonstrate the main trade-off in our model. The formal analysis is available

upon request. Consider a rm at the beginning of the rst period in a two-period risk-neutral economy. The sequence of events is

depicted inFig. 1 and numerical analysis is presented in Table 1. The rm has a protable investment project to undertake

immediately at Date 0. If undertaken, therm has to invest $50 and the project will generate a xed cash ow of $55 tothe rmat

the end ofthe rst period (Date 1). The rm also has a future investment project in the second period. However, whether it will be

protable or not is uncertain at the beginning of the rst period (Date 0). It is only at the end of the rst period (Date 1) that the

uncertainty will resolve. After therm learns about the expected NPV of the project at Date 1, it then decides whether to invest $55

in the second period project or not. For simplicity, we assume that the risk-free rate is zero.

We assume that the manager of the rm tries to maximize the rm's equity value (for example, the manager is the owner of the

rm). For simplicity, we assumethat even though all agents in the economy observe the expected NPVof the secondperiod project

when the uncertainty resolves at Date 1, the project NPV is not contractible. More specically, a contract that requires the manager

to invest in the second period project only if it has a positive NPV cannot be enforced by a court. 4 The project can be in three

possible states at Date 1: bad, mediocre, and good. The second period project is risky because it can yield a cashow of either $100

or $0. If the second period project is bad, it will yield a cashow of $100 with probability of 0.2, and a $0 cashow with probability

0.8. If it is mediocre, the probability of a $100 cashowis 0.7. If it is good, the probability ofa $100 cashow is 0.9. Noticethat it is

not efcient for the rm to invest if the project is bad because its NPV is $35.

None of the agents in the economy knows the state of the second period project at Date 0; however, they have a belief about it.They believe that the probability of a bad project is 0.25, a mediocre project is 0.5, and a good project is 0.25.

We focus on two debt nancing contracts of thermto nance the initial $50 investment: a non-callable bond maturing at the

end of the second period, or a callable bond with the same maturity but is callable at the end of the rst period.5 Neither pays any

coupon. We assume that the bond investors are competitive and require their investment to at least break even. We further

assume that each bond holder only buys a small fraction of the bond issued. As a result, non-callable bonds cannot be bought back

at Date 1 because of a hold-up problem (Gertner and Scharfstein, 1991). We will discuss this problem in more details later.

There are two factors determining therm's nancing choices: bond issuing cost and the risk-shifting cost. Bond issuing cost is

assumed to be a xed fee whenever the rm issues a new bond. That is, if the rm calls back its bond and issues a new bond to

invest in the second period, it has to incur another issuing cost. We also call the issuing cost of a new bond to renance the old one

arefunding cost. Clearly, if the rm issues a non-callable bond, it will never incur the refunding cost. Risk-shifting cost will be the

negative NPV incurred by the rm due to equity holders' risk-shifting incentive.

4

This is the common assumption in the incomplete contract literature (Grossman and Hart, 1986).5 We implicitly assume that the rm will issue debt rather than equity because of the benet of the debt, e.g., tax benet. See footnote 10 for more detail.

590 Z. Chen et al. / Journal of Corporate Finance 16 (2010) 588607
http://-/?-http://-/?-http://-/?-http://-/?-


4/20


5/20


6/20

bond investors learn about the project, the bond will be fairly priced. As a result, the gain to the equity holders is the project NPV

minus the refunding cost. The rm would thus invest in the project only if it is either good or mediocre. However, in these two

states, the rm has to reissue another short-term bond to nance the second period project at an issuing cost of $4. Comparing this

alternative with the callable bond, we can see that the callable bond dominates because it eliminates the risk-shifting cost while

the rm needs to incur an issuing cost only when the project is good.11

3.2. Testable hypotheses

Based on our analysis, we offer the following hypotheses.

H1. A rm expecting poorer future investment opportunities is more likely to issue a callable bond.

H2. A rm with higher leverage is subject to greater risk-shifting problems, thus is more likely to issue a callable bond.

H3. A rm with greater investment risk is more likely to issue a callable bond.

H4. Conditional on calling a bond, a rm with poorer performance is less likely to refund.

H5. Conditional on calling a bond, a rm with less active investments is less likely to refund.

H6. Arm with the best performance and most active investments tends to call and refund its bonds; a rm with the poorest

performance and least active investments tends to call without refund; a rm with mediocre performance and investments tendsto not call at all.

H2 and H3are not unique to our model. Both theories of solving debt overhang and reducing risk shifting suggest a positive

relation between leverage and the likelihood of issuing callable bonds. The signaling theory suggests that rms suffering from

more severe asymmetric information (e.g.,rms with greater investment risk) are more likely to issue callable bonds.H1, H4, H5,

and H6are unique to our model since none of the existing theories offer the same empirical implications. For example, the

signaling theory predicts that rms with better private information about future performance are more likely to issue callable

bonds. The solving debt overhang theory predicts that rms expecting better future investment opportunities would incur higher

costs of forgone investment due to debt overhang, thus would be more likely to issue callable bonds. The theory of removing

restrictive covenants predicts that rms with better investment opportunities should be more likely to issue callable bonds

because they would value the option to remove the restrictive covenants more. These theories all predict a positive relation

between a rm's future investment opportunities and its likelihood of issuing a callable bond, which is the opposite ofH1.

Furthermore, the theory of reducing risk shifting inBarnea et al. (1980)is silent regarding a rm's refunding decision upon their

calls, and other existing theories suggest that a rm should always refund its call. In contrast, H4 and H5predict when a rmshould refund its calls and when it shouldn't. H6predicts when a rm would call with refund, when it would call without refund,

and when it would not call at all. These four hypotheses help differentiate our theory from the others.

4. Data, variable construction, and descriptive statistics

4.1. Our sample of bonds

To investigate a rm's decision of issuing callable versus non-callable bonds, we obtain data on 13,784 nonconvertible xed

rate U.S. corporate bonds issued between January 1980 and December 2003 from the Fixed Investment Securities Database (FISD).

The FISD database (which is provided by LDS Global Information Services, Inc., currently owned by Mergent) contains issue- and

issuer-specic information, such as coupon rate, maturity, and credit rating, on all U.S. corporate bonds maturing in 1990 or later.

We use the FISD database instead of the New Issue Database of Securities Data Company (SDC) as the FISD database speci

esbonds as callable or non-callable, while the SDC database does not. More importantly, we nd that using information in the SDC

database to infer whether a bond is callable or non-callable may not lead to accurate categorization. 12

11 Another alternative is all-equity nancing. We rule out this alternative because of the benets of debt nancing, e.g., tax benets, which are not modeled in

our setting. However, if we consider the tax benets of debt nancing, we have to consider the extra cost of the callable bond because when the rm calls back

the bond without refunding, the rm loses its tax benet. The loss of tax benets will make the callable bond less favorable. But as long as the loss is not too big

relative to the risk-shifting cost, which is more likely to be the case when the rm does not have many good investment opportunities, it is still optimal for the

rm to issue a callable bond than a non-callable bond or equity nancing. The details are available from the authors.12 For example,Guntay et al. (2004)classify their bond sample into callable and non-callable bonds by examining the difference between the call protection

period and time to maturity. They dene a bond as being callable if the call protection period is less than one year, ve years, seven years, or ten years as the bond

will mature respectively within three to seven years, seven to ten years, ten to fteen years, or more than fteen years. To examine the validity of this

classication, we take all the nonconvertible xed rate bonds in the FISD being called up to year 2004, and hand match them to the bond issues from the SDC

database based on issuer cusip, issuer name, issuance date, maturity dates, and coupon rate. Thirty- ve percent of these bonds are actually being categorized as

non-callable bonds according to the above classication scheme. In contrast, only 1% of the bonds being called in FISD are misclassied as non-called bonds in

FISD. Therefore, we believe the denition of callable bonds in FISD is more reliable than the approximate classication based on the call protection period andtime to maturity in the SDC database.

http://-/?-http://-/?-http://-/?-http://-/?-


7/20

Fig. 2presents the distribution of all the 13,784 nonconvertible xed rate U.S. corporate bonds over time. Callable bonds were

very popular debt instruments in the 1980s, accounting for on average 70% of total public debt. The proportion of callable bonds

drops signicantly to only 23% in 1990s and early 2000s. The transition from high to low usage of callable bonds in early 1990s

accompanies a rapid growth in bond issuance, as shown inFig. 2. On average, callable bonds constitute 42% of total public debt

issued between 1980 and 2003. We also plot market interest rate (10-year Treasury rate) in Fig. 2.The signicant drop in the

percentage of callable bond issuances is coincident with the decreasing interest rate in our sample period. This evidence is

consistent with the theory of hedging interest rate risk.

After we merge the bond sample with the CRSP/Compustat database as well as excluded those bonds issued by rms without

operating income beta, we lost 4554 bonds.13 Utility and nancialrms often use callable bonds to hedge interest rate exposure

due to their duration gaps; however, they may not be appropriate targets for our study as we evaluate our theory on hedging

investment risk. Thus we exclude 3556 bonds issued by utility rms (SIC Codes between 4800 and 4999) and nancialrms (4-

digit SIC Codes between 6000 and 6999). As a result, our sample is reduced to 5674 bonds. To be included in our nal analysis, we

require a bond with complete issue-specic information, (e.g., issue amount and S&P credit rating). Furthermore, a bond issuer

must have stock prices available in the CRSP database and relevant accounting information available in the Compustat database

(e.g., total assets and long-term debt). This yields a nal sample of 3156 bonds issued between 1980 and 2003.

4.2. Variable construction for callable and non-callable bonds

4.2.1. Measuring future investment opportunities

We adopt several ex ante proxy variables to measure an issuing rm's future investment opportunities; the market/book ratio

(MB), the price/earnings ratio (PE), return on assets (ROA), analyst earnings forecasts (FORECAST), and growth rate of investment

( CAPEX and CAPEXRD).Unless otherwise noted, all variables are measured as of the year ending just prior to the bond issuance

date. Variable denitions are in the Appendix. Hypothesis H1 suggests that the probability of a rm issuing callable bond would benegatively related to these proxy variables of future investment opportunities. In contrast, theory of signaling, solving debt

overhang, and removing restrictive covenants all predict that the likelihood of issuing callable bonds would be positively related to

future investment opportunities.

4.2.2. Measuring leverage

Leverage is measured as the book value of either long-term debt or total debt (long-term debt plus debt in current liabilities)

divided by the book value of total assets. Our choice of book (rather than market) leverage is inuenced byWelch (2004), who

points out that market leverage may change passively simply because of changes in stock price performance.14 HypothesisH2

suggests that the probability of a rm issuing callable bond would be positively related to leverage.

Fig. 2. Bond sample distribution over time. The sample consists of 13,784 nonconvertible xed rate U.S. corporate bonds issued between January 1980 and

December2003 obtained from theFixedIncome Securities Database (FISD). We report thepercentage of callable bonds,the ten-year Treasury rate (obtained from

the constant-maturity Treasury bond yield from the H.15 release of the Federal Reserve System) ateach year-end, and the number of bonds issued every year.

13

About 2000 of them are lost due to the fact that some rms do not have sufcient quarterly data to compute operating income beta.14 Using market leverage ratio yields similar results.

http://-/?-http://-/?-http://localhost/var/www/apps/conversion/tmp/scratch_3/image%20of%20Fig.%E0%B2%80http://-/?-http://-/?-


8/20

Table 2

Summary statistics. Our sample includes 3156 bonds issued between 1980 and 2003. In Panel A, we report sample distribution in each one-digit SIC coded

industry. In Panel B, we report the sample distribution over time. In Panel C, we provide descriptive statistics on the issue-specic and rm-specic variables. All

variables are winsorized at the 1st and 99th percentile. Call dummy is a binary variable that equals one for callable bonds and zero for non-callable bonds. Issue

amount is thedollarproceeds of each bond issue. First-timeissuerdummy equals oneif this is therst time for a rmto issue a bondin the USpublicbondmarket

since January of 1975, and zero otherwise. Time to maturity is measured as the logarithm of the difference in years between the issuance date and maturity date.

Rating is thescoreof S&Prating, which is computed using a conversion process in which AAA+-rated bonds areassigned a value of 23 andD-rated bonds receive a

value of 1. Leverage is measured as the book value of either long-term debt or total debt (long-term debt plus debt in current liabilities) divided by the book value

of total assets. Firm size is dened as the logarithm of total assets. The market/book ratio (MB) is dened as the market value of total assets (sum of book value of

debt and market value of equity) divided by the book of value of equity. The price/earnings ratio (PE) is dened as stock price divided by earnings per share. ROA1is the ratio of operating income before interest, tax, and depreciation (EBITD) and the book value of total assets. ROA2 is net income scaled by the book value of

total assets. FORECAST1is themedian value of themost recentannual earnings forecasts forthe forthcoming scal year-end provided by all analysts. FORECAST2is

the median value of the most recent annual earnings forecasts for the scal year-end of next year provided by all analysts. Both FORECAST1 and FORECAST2 are

scaled by the year-end book value of equity. CAPEX (CAPEXRD) are the rst difference of capital expenditures (capital expenditures plus R&D expenses) scaled

by total sales. Risk-free rate is the constant-maturity Treasury bond yield from the H.15 release of the Federal Reserve System matching the maturity of each bond

issue. If the maturity of a corporate bond does not match that of a Treasury bond, we linearly interpolate theTreasury rates for maturities of one, three, ve, seven,

ten, twenty, and thirty years. Operating income beta is the slope coefcient from a regression in which we regress the quarterly changes in operating income

beforedepreciation normalized by total assetsover thelast 7 years precedingthe debt issue on changes in 1-yearT-bill rates.Operating incomevolatility is dened

as the standard deviation of the rst difference in quarterly earnings before interest, depreciation, and tax over the last 7 years preceding the debt issue,

normalized by theaverage value of total assetsover thesame time period. Unlessotherwisenoted, all variables aremeasuredas of theyear endingjust prior to the

bond issuance date.

Panel A. Sample distribution over industry

One-digit SIC Code Industry NOBS

0 Agriculture, forestry, and shing 61 Mining 269

2 Construction 955

3 Manufacturing 728

4 Transportation 534

5 Wholesale Trade 408

7 Agricultural Services 167

8 Forestry 89

Panel B. Sample distribution over time

Year NOBS % callable bonds (in terms of # of bonds) % callable bonds (in terms of issue amount)

1980 39 1.000 1.000

1981 26 1.000 1.000

1982 48 0.854 0.830

1983 32 0.938 0.979

1984 38 0.895 0.8851985 92 0.793 0.787

1986 166 0.608 0.689

1987 188 0.277 0.536

1988 93 0.591 0.750

1989 162 0.142 0.318

1990 154 0.026 0.037

1991 170 0.100 0.109

1992 216 0.269 0.244

1993 245 0.347 0.340

1994 101 0.297 0.399

1995 140 0.179 0.187

1996 177 0.266 0.235

1997 197 0.198 0.162

1998 223 0.152 0.140

1999 156 0.179 0.131

2000 93 0.075 0.0572001 148 0.108 0.054

2002 118 0.127 0.104

2003 134 0.187 0.155

Panel C. Descriptive statistics

Variable NOBS Mean Std. dev. Minimum Maximum

Call dummy 3156 0.2864 0.4522 0.0000 1.0000

Issue amount ($ million) 3156 195.32 162.31 0.37 1000.00

First-time issue dummy 3156 0.1518 0.3589 0.0000 1.0000

Time to maturity 3156 13.3279 8.4665 2.0164 40.0329

Rating 3156 13.5612 3.3910 1.0000 22.0000

Total assets ($ million) 3156 7960.69 9466.95 78.95 70349.00

Total market value ($ million) 3074 11955.67 15836.14 105.36 114839.09

(continued on next page)



9/20

4.2.3. Measuring investment risk

We employ several variables to proxy for investment risk, including rmsize, a rst-time issuer dummy, and operating income

volatility. Smaller rms, rst-time issuers, and rms with larger operating income volatility would have greater level of

investment risk. Our hypothesisH3suggests that the probability of a rm issuing a callable bond would be negatively related to

rm size, but positively related to the rst-time issuer dummy and operating income volatility.

A rm's investment risk or overall risk is also reected in bond credit rating. RATING is the S&P credit rating score, which is

computed using a conversion process in which AAA+-rated bonds are assigned a value of 23 and D-rated bonds receive a value of

1. Since credit rating may incorporate part or all of future investment opportunities, we orthogonalize this variable by regressing it

against eachof the variables proxied for investment opportunities, and use the residual term (Rating Residual) as the regressor in

the model.15 HypothesisH3suggests that the probability of a rm issuing a callable bond would be negatively related to Rating

Residual.

4.2.4. Other control variables

Guntay et al. (2004)show that the choice of issuing a callable bond is positively related to the market interest rate and a rm's

interest rate sensitivity of operating income. Based on this evidence, they argue that a rm uses a callable bond to hedge operating

income uctuations. To control for the confounding effects of market interest rate and a rm's operating income exposure to

interest rate, we also include the risk-free rate and the operating income beta, which measures a rm's interest rate sensitivity to

operating income.

We also include a few issue-specic variables that might affect the choice of whether to issue a callable or a non-callable bond,

including time to maturity and issue size. According to the theory of hedging interest rate risk, there is a substitution effect

between using a call option and shortening maturity. Thus we expected a positive relation between maturity and the probability of

issuing callable bonds. In addition, larger issues are more likely to be associated with callable bonds since they create higher

interest rate exposure for a rm. It is worth mentioning that our theory also suggests a positive relation between issue size and

maturity, and the probability of issuing callable bonds. This is because larger issues and longer maturities would subject the issuers

to greater investment risk.

5. Choices of issuing a callable bond

Table 2reports descriptive statistics for our nal bond sample.16 Panel A presents sample distribution in each one-digit SIC

coded industry; Panel B presents sample distribution over time; Panel C offers summary statistics on the variables used in theanalysis. The bond sample contains about 29% callable bonds;17 15% are rst-time issue. The average issue size is $195 million,

while the average time to maturity is approximately 13 years, suggesting a large proportion of long-term bonds in the sample. The

average S&P credit rating score is 13.6, equivalent to a rating between BBB+ and BBB. Our sample seems to be lled with large

companies. The average total asset of bond issuers is $7.9 billion, and their average market value is about $12 billion.

5.1. Univariate results

InTable 3, we examine the difference in mean and median value of issue-specic andrm-specic variables between callable

and non-callable bond issues. We observe signicant differences in both mean and median values of proxy variables of future

investment opportunities between the two groups. Callable bonds are issued by rms with a lower market/book ratio (MB), lower

Table 2 (continued)

Panel C. Descriptive statistics

Variable NOBS Mean Std. dev. Minimum Maximum

Firm size (Ln Assets) 3156 8.3483 1.2368 4.3688 11.1612

Leverage (long-term debt) 3134 0.2679 0.1339 0.0084 0.7945

Leverage (total debt) 3156 0.3128 0.1349 0.0382 0.8450

PE 3068 15.4538 21.8513 112.5000 227.5735

MB 3035 1.4796 0.6257 0.8141 4.5558FORECAST1 2688 0.0976 0.1804 2.0239 1.7613

FORECAST2 2649 0.1216 0.1758 2.0239 1.7613

ROA1 3102 0.1461 0.0547 0.0049 0.3078

ROA2 3090 0.0439 0.0429 0.1847 0.1679

CAPEX 3040 0.0069 0.0489 0.3576 0.3052

CAPEXRD 3040 0.0071 0.0502 0.3576 0.3208

Risk-free rate 3156 7.0083 1.9826 1.6682 15.1599

Operating income beta 3123 0.0007 0.0266 0.1345 0.1376

Operating income volatility 3156 0.0137 0.0090 0.0023 0.0616

15 The residual term from the regression captures the credit rating information without the in uence of investment opportunities.16

To minimize the effect of outliers, we winsorize all the variables at the 1st and 99th percentiles.17 Callable bonds account for 24% of the total issue amount in our sample.

http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-


10/20

price/earnings ratio (PE), lower ROA, and lower analyst forecasts for future earnings. Growth rate in capital expenditure and R&D

expenses is lower in rms issuing callable bonds than that in rms issuing non-callable bonds. These results support hypothesis

H1: Firms with poorer future investment opportunities are more likely to issue callable bonds. Furthermore, callable bond issuers

have greater mean and median values of leverage, supporting hypothesis H2. Callable bonds are issued by smaller rms with lower

credit ratings and greater operating income volatility, and are more likely to be rst-time issues. These results are consistent with

hypothesisH3: Firms with greater investment risk are more likely to issue callable bonds. Both types of bond issuers have a very

small mean or median operating income beta; although the mean is not statistically signicantly different between the two

groups, the median operating income beta of callable bond issuers is signicantly higher than that of non-callable bond issuers.

Consistent with the theory of hedging interest rate risk, callable bond issuances are associated with a signi cantly higher interest

rate. In addition, we nd callable bond issuances are associated with longer maturity and smaller issue size.

5.2. Logistic regressions explaining the likelihood of issuing callable bonds

We employ logistic regressions to explore the cross-sectional relation between a rm's likelihood of issuing a callable bond and

variables that proxy for future investment opportunities, leverage, and investment risk. The dependent variable in the logistic

models is a binary variable equal to one for callable bonds and zero for non-callable bonds. The results are reported in Table 4.18

As shown inTable 4, the explanatory power of our logit models is substantial, as evidenced by the Pseudo-R2 exceeding 62% in

each regression. The rst variable of interest is market/book ratio (MB), which proxies for future investment opportunities. The

coefcient estimate on MB is negative and statistically signicant at the 1% level in all models. This result is consistent with

hypothesisH1,suggesting that rms with better future investment opportunities are less likely to issue callable bonds.19

Our theoretical analysis indicates that callable bonds could resolve the agency problem of risk shifting when a

rm's futureinvestment opportunities are poor. Hypothesis H2 suggests that a rm with a higher leverage ratio is more likely to issue a callable

bond, since it is subject to a greater debt agency problem. Consistent with H2, we observe a positive and signicant relation

between the total leverage ratio and the probability of issuing a callable bond in model (1). To test the robustness of this leverage

effect, we include in model (2) a long-term leverage ratio, and the result remains. 20

We include several variables to proxy for investment risk. Firm size is signi cantly negatively related to the probability of

issuing a callable bond, since a larger rm is often subject to less investment risk. First-time issuers tend to be smaller rms, or

rms with less experience and reputation (or access) in the public debt market. The coefcient estimate of the rst-time issuer

dummy is positive and signicant. Operating income volatility, however, is not signicantly related to the usage of a callable bond.

Rating residual is negatively related to the probability of issuing a callable bond, and the coefcient estimate is highly signicant. A

rm with a higher credit rating residual is facing lower investment risk, and hence, it is less likely to issue a callable bond .These

results support hypothesis H3: a rm with greater investment risk is less likely to issue a callable bond.Kish and Livington (1992)

and Crabbe and Helwege (1994)document a signicant negative relation between credit rating and the use of callable bonds.

To assess the economic impact of each variable on the choice of issuing callable bonds, we compute an odds ratio thatrepresents the change in probability of issuing a callable bond given the change of one standard deviation of each independent

variable. Change in probability for MB is 0.3564 in model (1), implying that an increase of one standard deviation in MB would

decrease the probability of issuing a callable bond by 36%.21 Change in probability for total leverage ratio and rating residual is

0.2388 and 0.5594, respectively. These results suggest an economically signicant effect of future investment opportunities,

leverage, and rating residual on the likelihood of issuing a callable bond.

The coefcient estimate of the risk-free rate is positive and signicant at the 1% level in most regressions, suggesting that

interest rate risk may be a signicant consideration in corporate usage of callable bonds.22 Guntay et al. (2004)argue that if

callable bonds are used for hedging interest rate risk, rms with higher interest rate sensitivity (operating income beta) would be

more likely to issue callable bonds. Our evidence does not support their argument. We nd thatthe coefcient estimates on

operating income beta are mostly positive; however, they are not signicant in any of the regressions.23 Overall, our analysis offers

mixed evidence with respect to the theory of hedging interest rate.24

18

To control for time and industry effects, we include in the logistic regressions dummy variables for each calendar year and each industry based on two-digitSIC Codes.19 One might argue that MB might capture the risk aspect of a rm since a rm with high growth potential would have a large MB but also a high level of

investment risk. In our logistic models, we include several variables to control for investment risk as discussed above; therefore, the relation we observe between

MB and the probability of issuing a callable bond should reect the impact of future investment opportunities rather than investment risk on the usage of a

callable bond. Furthermore, since the relationship between investment risk and callable bond usage is expected to be positive, the negative relation between MB

and the probability of issuing a callable bond is likely the impact of future investment opportunities that is captured by MB.20 Kish and Livington (1992)have documented a similar result.21 Given that the mean probability of issuing a callable bond is 14.31%, as indicated in model (1) in Table 4, this is equivalent to an increase of unconditional

probability of issuing a callable bond by 5.2%.22 This result is consistent with the ndings documented in the literature (e.g., Kish and Livington, 1992:Guntay et al., 2004).23 To take into account the statistical signicance of the beta estimate, as inGraham and Rogers (2002), we dene an operating income exposure variable that is

zero if the operating income beta is not signicant at the 10% level. Otherwise, it takes the value of 1 or +1, depending on the sign of the coefcient. Our

results are similar as we use the operating income exposure variable in the analysis.24 To assess whether the difference between our results on operating income beta and those inGuntay et al. (2004)is driven by different sample periods, we

estimate the same regression models inTable 3ofGuntay et al. (2004)based on a sample period of 1981 through 1997. The coefcient estimate of operating

income beta remains insignicant. Nevertheless, the difference between our results and those inGuntay et al. (2004)might be driven by the use of differentdatabases and/or different methods of dening a callable bond, as discussed in footnote 10.

http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-


11/20

As withKish and Livingston (1992), Crabbe and Helwege (1994), and Guntay et al. (2004), we nd that larger bond issues and

those with longer maturities are more likely to be callable. In model (3), we replace time to maturity with duration, that is, the

discount time-weighted cash ows of the bond divided by bond price, and obtain similar results. The coefcient estimate on

duration is signicantly positive. This nding is consistent with the theory of hedging interest rate risk. Since larger bond issues

and issues with longer maturities are associated with greater interest rate risk, the issuing rm is more likely to use a callable bond

to hedge interest rate risk. This nding is also consistent with our theory that longer maturity and larger issue size may be

associated with greater future investment risk. Therefore, a rm would be more likely to issue a callable bond to minimize the

agency problem according toH3.

Our logistic analysis above focuses on a rm that issues callable and non-callable bonds; however, the decision of whether or

not to issue a bond could itself be endogenous. Hence, we investigate the possibility that our results are spuriously driven by an

unobserved but nonrandom selection criterion. To test (and if necessarily correct) for selection bias, we estimate a maximum

likelihood version of aHeckman (1979) selection model of regression, and the result is reported in model (4) ofTable 4. In

particular, we take all the rm-year observations in the Compustat database in 1980 through 2003, and construct a dummy

variable issuing bondbased on whether or not arm issued a bond (callable or non-callable) in a particular year as recorded in

the FISD database. Then we estimate a selection probability model (results not reported) that relates the probability of issuing

bonds to rm size, MB, R&D expenses (normalized by assets), proportion of tangible assets, leverage, ROA, Altman z-score, and

operating income volatility.25 These variables are chosen based onDenis and Mihov (2003). As shown in model (4), the inverse

Mills ratio from the selection probability model is only marginally signicant at the 10% level. After controlling for potential

selection bias, our results are robust. The selection-adjusted coefcient estimates are similar to those from model (1).

In addition, to assess whether our results are driven by rm size, we divide our sample into three equal groups based on size

(small, medium, and large), and a logistic regression is conducted in each sub-sample. Our results are robust in each group. The

coefcient estimates on MB (measure of future investment opportunities) are signicantly negative in all three sub-samples, and

the magnitudes of coefcients are also comparable among groups (results available upon request).

5.3. Additional tests of hypothesisH1

To further investigate the relation between future investment opportunities and the probability of issuing callable bonds (H1),

we employ several alternative ex ante measures of investment opportunities; however, regressions results are not reported but

available upon request. The coefcient estimates of price/earnings ratio (PE), ROA, and analyst earnings forecasts for the

forthcoming year and next year (FORECAST1 and FORECAST2) are all negative and statistically signicant at the 1% level. Changes

in probability indicate that an increase of one standard deviation in one of these variables is associated with a decrease of the

probability of issuing a callable bond by 10% to 32%. These results lend strong support for hypothesis H1.

To assess the relation between a rm's decision to issue a callable bond and its expected future investments, we also include an

ex ante measure of growth in capital expenditure (CAPEX) or growth in capital expenditure and R&D expenses (CAPEXRD). We

nd that the coefcient estimates ofCAPEX and CAPEXRD are both negative and signicant (results available upon request).

25

The dependent variable for the selection model is a binary variable

issuing bond

that equals one if a rm issued a callable or non-callable bond in a givenyear, as recorded in the FISD database, and zero otherwise.

Table 3

Univariate statistics of callable and non-callable bonds. This table reports the mean and median of issue-speci c andrm-specic variables for callable and non-

callable bonds issuedbetween 1980and 2003. T-tests (Wilcoxonrank tests)are used to examine thenull hypothesis that themean (median) of each variable is the

same between callable and non-callable bonds.

Variable Mean Median

Callable Non-callable T-statistics Callable Non-callable Z-statistics

MB 1.308 1.545 11.530 1.184 1.323 10.023

PE 11.520 16.986

6.280 11.275 15.218

10.884ROA1 0.144 0.147 1.120 0.146 0.146 0.494

ROA2 0.038 0.046 4.630 0.042 0.047 2.919

FORECAST1 0.058 0.110 5.050 0.053 0.077 11.346

FORECAST2 0.100 0.128 3.080 0.070 0.094 8.291

CAPEX 0.001 0.009 3.850 0.001 0.002 3.763

CAPEXRD 0.002 0.009 3.480 0.001 0.003 3.832

Leverage (total debt) 0.350 0.298 8.540 0.321 0.293 6.042

Firm size (Ln assets) 7.609 8.645 20.690 7.600 8.742 19.270

Rating 11.887 14.233 15.290 12.000 14.000 13.537

Operating income volatility 0.015 0.013 5.170 0.012 0.011 5.086

First-time issue dummy 0.254 0.111 9.030 0.000 0.000 10.240

Operating income beta 0.002 0.000 1.110 0.002 0.002 2.560

Risk-free rate 8.049 6.590 16.770 7.520 6.494 15.082

Time to maturity 15.742 12.359 9.680 10.025 10.016 8.001

Issue amount ($ million) 166.097 207.052 7.580 148.800 200.000 5.519

http://-/?-http://-/?-http://-/?-


12/20

Assuming that an ex ante growth of investment is a proxy of future investment growth, our results suggest that a rm expecting

high growth in investments would be less likely to use a callable bond, since it faces less investment risk in the future.

In addition to these ex ante measures of investment opportunities, we adopt a few ex post variables to proxy for investment

opportunities. Based on rational expectations, observed investment opportunities should be a proxy for anticipated investment

opportunities (e.g., Pilotte, 1992). We include a few ex post variables that proxy for investment opportunities: Ex-ROA is the

average ROA over the three years following bond issuance and Ex-CAPEX (Ex-CAPEXRD) is computed as the average ofCAPEX

(CAPEXRD) over the three years following bond issuance. The coefcient estimates of Ex-ROA, Ex-CAPEX, and Ex-CAPEXRD

are all negative and statistically signicant (results available upon request). These results further conrmthata rm is less likely to

issue a callable bond when it expects future investment opportunities to be better.

Our empirical evidence on the relation between future investment opportunities and the probability of issuing callable

bonds lends strong support to our hypotheses, particularlyH1. In contrast, this evidence is not consistent with the alternative

explanations, including the signaling theory, the theory of solving debt overhang, and removing restrictive covenants. As

discussed inSection 3, these three theories all predict that

rms with better future investment opportunities are more likely toissue callable bonds.

Table 4

Logistic regressions explaining issuance of callable bonds. This table presents the results of logistic models in which the dependent variable is a binary variable

equal to onefor callable bonds andzero fornon-callable bonds.Independent variables include issue-specic and rm-specic variables that proxyfor rms' future

investment opportunities, leverage, and investment risk. All the explanatoryvariables are as denedin the Appendix. Regression (4) controls forsample selection

bias by estimating a MLE version of theHeckman (1979)selection model. To control for time and industry effects, we also include dummy variables for each

calendar year andeach industry based on two-digitSIC Code. P-values arereportedin parentheses. Mean probabilityis thepredicted probabilityof issuingcallable

bonds when all explanatory variables have their mean values. Change in probabilityis denesas the percentage change in the probability of issuingcallable bonds

when the corresponding explanatory variable is increased by one STD, and all other variables are evaluated at their means and reported in { }.

Independent variables 1 2 3 4

Intercept 7.0718 7.1908 8.6560 8.1241

(b.0001) (b.0001) (b.0001) (0.0074)

MB 0.7946 0.7974 0.7372 0.5317

(b.0001) (b.0001) (b.0001) (b.0001)

{0.3564} {0.3579} {0.3376} {0.1971}

Leverage (total debt) 1.8859 1.3633 2.6990

(0.0002) (0.0186) (b.0001)

{0.2388} {0.1707} {0.1445}

Leverage (long-term debt) 2.0520

(0.0001)

{0.2598}

Firm size 0.5718 0.5590 0.4917 0.2052

(b.0001) (b.0001) (b.0001) (0.3288)

{0.4703} {0.4629} {0.4222} {0.149}

First-time issuer dummy 0.3672 0.4039 0.3261 0.1959

(0.0349) (0.0213) (0.0997) (0.2976)

{0.1188} {0.1316} {0.1063} {0.0244}

Rating residual 0.2928 0.2876 0.2318 0.3217

(b.0001) (b.0001) (b.0001) (b.0001)

{0.5594} {0.5532} {0.4773} {0.3682}

Operating income volatility 2.2343 3.0814 2.2243 3.9593

(0.7374) (0.6446) (0.7615) (0.4935)

{0.0178} {0.0246} {0.0179} {0.0412}

Risk-free rate 0.3567 0.3622 0.5813 0.0187

(b.0001) (b.0001) (b.0001) (0.8748)

{0.7721} {0.7892} {1.4651} {0.0129}

Operating income beta 0.4081 0.7013 1.5955 0.1488

(0.8476) (0.7417) (0.4924) (0.9132)

{0.0095} {0.0164} {0.0380} {0.0081}

Log(Issue amount) 0.6463 0.6443 0.7340 0.7313

(b

.0001) (b

.0001) (b

.0001) (b

.0001){1.2051} {1.2049} {1.4582} {0.3009}

Log(Time to maturity) 1.3484 1.3487 1.5264

(b.0001) (b.0001) (b.0001)

{0.9246} {0.9279} {0.25}

Duration 0.1185

(b.0001)

{0.6285}

Inverse Mills Ratio 0.9250

(0.0648)

Mean probability 0.1431 0.1416 0.1332 0.6693

Industry dummy & Calendar year dummy Yes Yes Yes Yes

Pseudo-R2 0.6321 0.6318 0.6297 0.6465

NOBS 3056 3038 2685 3056



13/20

5.4. Robustness tests on the choice of issuing callable bonds

One caveat of our results is that it does not take into account other nancial contracting devices, e.g., leverage and debt

maturity, which would also mitigate the risk-shifting incentive. While we have controlled for the impact of leverage and maturity

in our regression models above, the control might be problematic since the choice of a callable bond is likely jointly endogenous

with these controlvariables. As such, we conduct a few robustness tests to address the endogeneity issue of leverage and debt

maturity. 26

First we estimate a reduced form model that excludes Leverage and Log(Time to maturity) that are likely jointly endogenous,

and the results are reported in column (1) ofTable A1. As with the results in model (1) ofTable 4, the coefcient estimate on MB is

signicantly negative. Firm size and Rating residual are both signicantly negatively related to the probability of issuing a callable

bond.

Second we estimate a system of three simultaneous equations that recognizes that both leverage and debt maturity are

determined endogenously with the choice of issuing a callable bond. For the leverage and maturity equations in columns (3) and

(4) respectively, we use the explanatory variables thatJohnson (2003) and Billett et al. (2007)employ in their system of leverage

and maturity equations. In particular, in the leverage equation we include MB (market-to-book), operating income volatility, debt

maturity (prop. short-term debt), interaction term of MB and debt maturity, xed assets, protability, rm size, investment tax

credit dummy, net operating loss carry forward dummy, and abnormal earnings. In the maturity equation, we include MB

(market-to-book), leverage, operating income volatility, rm size and the square ofrm size, investment tax credit dummy, net

operating loss carry forward dummy, abnormal earnings, asset maturity, and rated rm dummy. In the equation explaining

the choice of issuing callable bonds (column 2), we include the same set of variables used in model (1) ofTable 4, except that

we include rm level debt maturity (prop. short-term debt) instead of bond maturity. Following Johnson (2003), maturity

(prop. short-term debt) is dened as the fractionof a rm's total debt that matures in 3 years orless. Wedene all other variablesin

Table A1. Each variable is measured at the scal year-end prior to the bond issuance date. The system of equations is estimated by

nonlinear two-stage least squares method for the pooled sample of callable and non-called bonds issued between 1980 and 2003.

After accounting for the endogenous choice of leverage and debt maturity, our results on the likelihood of issuing callable

bonds as reported in Table 4 remain robust. Wend that MB, rm size, and rating residual are all signicantly negatively related to

the probability of issuing a callable bond, which supports our hypothesis H1 and H3. Leverage remains signicantly positively

related to the likelihood of issuing a callable bond, supporting our hypothesis H2.27

6. Firms' choices of call with refund, call without refund, and not call

In addition to the implications on a rm's choice of whether to issue a callable or a non-callable bond, our model also provides

explicit empirical implications on a rm's choice whether to call the bond, as well as whether to refund the call. To test those

implications, we rst explore that, conditional on the call events, how a rm's performance and investment activity are related to

the decision on whether or not to refund the call. Second, we examine the impact ofrm performance and investment activity on

the rm's choices of call with refund, call without refund, and not call at all.

6.1. Choice of refunding around the call events

Our sample of bonds being called is obtained from the le called Amount Outstandingin the FISD.28 To be included in our

analysis, we require that a called bond have issue-specic information available in FISD (e.g., issue amount and credit rating) and

relevant accounting information available in the Compustat database (e.g., total assets, long-term debt) around the call date. The

analysis yields 853 bonds being called between 1983 and 2004.

To measure a rm's refunding activities around a call event, we aggregate the total amount of new debt within a 12-month

period surrounding the call date (i.e., six months before or six months after the call), as reported in the SDC New Issue Database.

New debt includes public debt, private placements, 144A, shelf registration debt, and convertible debt. We nd only 46% ofrms

issuing new debt in this 12-month window.

29

To de

ne a refunding call, we follow King and Mauer (2000) and require therefunding ratio (the total amount of new debt raised within a 12-month period divided by the book value of the bonds being

called) to be at least 110%.30 Otherwise, the call event is dened as a non-refunding call.

26 We thank an anonymous referee for suggesting this great point.27 The signs of the coefcient estimates on variables in the leverage and maturity equations are in general consistent with those reported inJohnson (2003) and

Billett et al. (2007).28 This le provides the date and amount of any changes to a bond issue's amount outstanding due to various actions (e.g., part of an issue called, entire issue

called, call with an equity clawback provision). This le allows us to identify those bond issues that are entirely called back up to 2004, including call date, call

price, and the amount of issue being called. We do not include bonds partially called or called with an equity clawback provision.29 King and Mauer (2000) report that 23% of their call events in 19751994 are associated with raising new debt. We nd a much higher percentage of

refunding calls in our sample, partly due to the fact that the SDC database offers a more comprehensive coverage of new debt issuance than Moody's manuals, the

Wall Street Journal Index, or LexisNexis thatKing and Mauer (2000)rely on to identify new nancing activities.30 The extra 10% in the refunding ratio could be thought of the residual nancing activities for an average rm in a given 12-month period. Even in the absence

of refunding calls, an average rm may raise some funds in the nancial market on a regular basis. We employ alternative denition of refunding calls (e.g., arefunding ratio of 100% or 130%), and our results remain qualitatively the same.

http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-


14/20

Panel A inTable 5reports descriptive statistics on our bond sample being called from 1983 to 2004. Despite the fact that only

46% ofrms raise new debt within a 12-month period surrounding the call date, these rms raise about $356 million in new debt

on average, which is equivalent to 3.07 times the average amount of bonds being called back. Forty percent of the call events arecategorized as refunding calls since they raise new debt at least as large as 110% of the book value of the bonds. Therefore, there are

a signicant proportion ofrms (60%) not refunding their calls.

In Panel B ofTable 5, we investigate the difference between refunding calls and non-refunding calls. For those refunding calls,

the mean refunding ratio is 7.52, suggesting that many rms raised a large amount of capital within a 12-month period around the

call date. Consistent with our sample construction, non-refunding calls are associated with little new debt. The mean refunding

ratio is 0.075 in this group, suggesting that non-refundingrms raise on average new capital of only 7.5% of the amount of bonds

being called. This raises an interesting question: Why are these rm calling back bonds without refunding? One might argue that it

is not worth refunding because the interest rate does not drop enough to outweigh the refunding costs. If this is true, we would

expect the change of interest rate between call date and issue date to be larger (or less negative) in the non-refunding group than

that in the refunding group. As shown in Panel B ofTable 5, the mean change of interest rate between the call date and issue the

date is negative in both groups; a t-test indicates that the mean change is not statistically signicantly different, suggesting that an

interest rate change is not the reason for a rm to not refund a call.

In Panel B ofTable 5,we also examine the credit rating of these two groups of bonds at the time when they were issued andcalled. When they were issued, bonds in the refunding group (average rating of BBB+) were rated two notches higher than those

in the non-refunding group (average rating of BBB); upon their calls, they basically retain their rating level. These ndings are

not consistent with the theories of signaling and solving the debt overhang problem, since both predict an improvement in a rm's

prospects upon calling a bond (i.e., an improvement in credit rating).31 In addition, the average number of years between the call

date and the maturity date is 7.70 and 9.37 years for the non-refunding and refunding group respectively. This suggests that the

call events in our sample are signicant early terminations of bond maturity.

While we count both public and private debt in measuring refunding activities, we do not include bank debt. If the rms in our

sample refund public debt with bank debt, we may misclassify refunding calls as non-refunding. To address this issue, we examine

the change of total debt in the balance sheet from before to after the call event. We nd that a non-refunding rm experiences a

signicant decrease in total leverage (total debt divided by assets), while the refunding rm has a signicant increase in leverage

Table 5

Summary statistics on called bond sample. Our sample includes 853 bonds being called back between 1983 and 2004. In Panel A, we provide descriptive statistics

on the called bond sample. Total new debt is the total amounts of new debt (public and private) issued within a 12-month period surrounding the call dates.

Refunding ratio is total amounts of new debt raised within a 12-month period divided by the book value of the bonds being called. Refund dummy is a binary

variable equal to one if the refunding ratio is 110% or greater. Otherwise, the refund dummy is zero. Tangible assets are dened as property, plant and equipment

divided by total assets. All other variables are as dened in the Appendix. In Panel B, we report for both refunding calls and non-refunding calls, the mean of

refunding ratio, change of interest rate (maturity matched Treasury rate at call date minus Treasury rate at issue date), credit rating at issue date, credit rating at

call date, the change of credit ratings between call date and issue date, and the number of years from the call date to the maturity date.T-tests are conducted to

examine the null hypothesis that the mean of these two groups is the same.

Panel A. Descriptive statistics

Variable NOBS Mean Std dev Minimum Maximum

Total new debt 853 356.2121 607.3295 0.0000 4329.2000

Refunding ratio 853 3.0696 6.1880 0.0000 64.0000

Refund dummy 853 0.4021 0.4906 0.0000 1.0000

ROA1 853 0.1377 0.0551 0.0075 0.2796

ROA2 833 0.0301 0.0475 0.1702 0.1379

CAPEX 814 0.0048 0.0410 0.2525 0.2066

CAPEXRD 814 0.0048 0.0420 0.2525 0.2066

Firm size 853 8.2935 1.3425 4.7089 11.0973

Tangible assets 853 0.4377 0.2187 0.0213 0.9053

Operating income volatility 853 0.0134 0.0088 0.0023 0.0612

Leverage 853 0.3601 0.1558 0.0610 0.8886

Panel B. Mean difference between refunding calls and non-refunding calls

NOBS Refunding

ratio

Change of

risk-free rate

Rating

at issue

Rating

at call

Rating at

callRating at issue

Years from maturity

date to call date

Non-refunding calls 510 0.0752 2.1692 12.2176 12.2294 0.0118 7.6789

Refunding calls 343 7.5219 2.3172 14.3907 14.3703 0.0204 9.4608

Difference 7.4467 0.1480 2.1731 2.1409 0.0322 1.7819

T-statistics 17.50 1.11 8.28 8.39 0.31 3.35

31

Crabbe and Helwege (1994) and King and Mauer (2000)also document little signicant rating improvement for a sample of bonds being called in 1975

1994.

http://-/?-http://-/?-


15/20

(results available upon request). Further analysis indicates that the change in total leverage largely comes from long-term rather

than short-term debt, indicating that our denition of refunding is accurate.

As we show in our model, as a rm faces deteriorating investment opportunities, it would choose to not invest in the project

and call back a bond without refunding. In contrast, when a rm has excellent investment opportunities, it would invest, call back

the bond, and refund the call. Hypotheses H4 and H5suggest that a rm with poorer performance and less active investments is

less likely to refund when it calls a bond.

To test these two hypotheses, we employ logistic models to investigate the cross-sectional relation between a rm's likelihood

of refunding and its performance and investment activities. The dependent variable in the logistic models is a binary variable equal

to one for refunding a call and zero for not refunding a call. Explanatory variables include a measure of performance, ROA1 (EBITD/

Assets) or ROA2 (net income/assets), or a measure of investment activities, CAPEX or CAPEXRD. In addition, we include control

variables that proxy for nancing costs, includingrm size, tangible assets (dened as property, plant, and equipment divided by

total assets), total leverage ratio, and operating income volatility. All these explanatory variables are computed in the year-end

prior to the call date. Furthermore, we include the change of interest rate and the change of credit rating between the call date and

issue date to capture the potential benet of refunding.32

The regression results are reported inTable 6. The variables of interest are the measures of performance and investment

activities. The coefcient estimates on ROA1 and ROA2 are both positive and statistically signicant at the 5% level, suggesting that

a rm with poorer performance is less likely to refund a call. Furthermore, the growth rates of CAPEX and CAPEXRD are both

signicantly positively related to the probability of refunding around a call event. The evidence indicates that a poor-performing

rm invests less in their project and pays out cash. Our results are also economically signicant. A decrease of one standard

deviation in ROA2 and the growth rates of CAPEX are associated with a decrease of 26% and 12% in the likelihood of refunding,

respectively. These results support hypothesesH4 and H5.

The coefcient estimate ofrm size is signicantly positive in all the regressions, suggesting that a larger rm is more likely to

refund a call since it incurs fewernancing costs due to its better access to the capital market and the economy of scale effect. The

proportion of tangible assets is positively related to the probability of refunding around call events; however, the coefcient

estimates are statistically insignicant. The coefcient estimates of total leverage ratio and operating income volatility are not

statistically signicant either. The change of interest rate between the call date and the issue date is not signicantly related to the

probability of refunding. This result is due to the inclusion of dummy variables for each calendar year. If we leave out calendar year

dummy variables, the coefcient estimate on the change of interest rate becomes negative and statistically signicant. This nding

suggests that the lower the interest rate at the call date, the more likely a rm is to refund a call because of the benets of

refunding.33 Change in credit rating (rating at call minus rating at issue) is not signicantly related to the choice of refunding. This

evidence is inconsistent with the theory of signaling and solving debt overhang problem.

6.2. Robustness tests

We conduct the following robustness checks of refunding choice. First, we restrict our sample to called bonds that are also

present in our analysis of the choices of issuing callable versus non-callable bonds, and this yields 492 called bonds. 34 The logistic

regression results are similar to those using the full bond sample, as reported in Table 6. Second, most of previous studies on

callable bonds are based on call events hand collected from Moody's Manuals ( Vu, 1986; King and Mauer, 2000; Guntay et al.,

2004). As an alternatively sample source, we followKing and Mauer (2000)and consult Moody's Annual Bond Records, and hand

collect 489 bonds called by industrial rms from 1990 to 2005. The results are similar.

Another, perhaps more extreme way for a rm to reduce investment than reducing capital expenditures is asset sales. H5

suggests that a non-refundingrm is more likely to sell assets around the call event. To test this hypothesis, we collect asset sales

activities from the SDC M&A database for the six months before and six months after each call event. We nd support forH5:the

average net asset sales (the amount of assets sold minus the amount of assets bought during the 12-months window) normalized

by the amount of debt being called are signicantly larger in the non-refunding group than that in the refunding group (results

available upon request).

6.3. Choice to call with refund, call without refund, or not call

We next examine a rm's unconditional choices to call with refund, call without refund, or not call as its bond exits the call

protection period. HypothesisH6 suggests that the choice is not monotonic with respect to rm performance and investment

activity. The best rms are more likely to call with refund. The worst rms are more likely to call without refund, while the

mediocrerms are more likely to not call their bonds.

We estimate a multinomial logit model to explore the impact of a rm's performance and investment activities on the three

choices of callable bonds: call with refund, call without refund, or not call. Following Denis et al. (1997), Shumway (2001), and

King and Mauer (2009),we start with a sample of all callable bonds in FISD. We track each callable bond starting from the year in

32 We include dummy variables for each calendar year and industry based on two-digit SIC Codes to control for the time and industry effects.33 The benet of refunding at a lower interest rate can be thought of as a deceased refunding cost in our model.34

There are 53% of bonds (492) being called in FISD also present in the at issue analysis in section 5; the rest are either callable bonds issued before 1980 orthey do not have adequate data available for issue analysis.

http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-


16/20

which its call protection expires or 1980 (whichever comes later) until the year of being called or 2004 (whichever comes rst).

Each bond in a year is categorized as either not called, called with refunding, or called without refunding.35 For example, if a bond

was called with refunding in 2002 but it became callable in 1995, we would have a time series of observations of the bond for 1995

through 2002. The bond would be categorized as not called in 1995 through 2001, and called with refunding in 2002.

Therefore, the choices in our study are both time series and cross-sectional.

An important caveat for our

not call

observations is that they might include

nancially distressed

rms that cannot afford tocall their bonds, and our model does not account for this scenario (in our model, the rm always has enough cash to call the bond

without refund if it chooses to). Failure to account for the rm's inability to call would potentially bias the performance and

investment activities of our not callsample downward. To mitigate this potential problem, we impose a simple selection lter.

We restrict our sample rms to those that are present in the CRSP database at the end of 2004.36

Table 7 reports the results of the multinomial logit models.37 We use the observations ofnot call as the base case and evaluate

the other two outcomes (call with refund and call without refund). Models (1), (3), (5), and (7) evaluate the choices of call with

refund and not call. Models (2), (4), (6), and (8) evaluate the choices of call without refund and not call. Independent variables

Table 6

Logistic regressions explainingthe likelihoodof refundingaround call events. This table presents theresults of logistic models in which thedependent variable is a

binary variable equal to one for refunding calls and zero for non-refunding calls. Independent variables include rm size, leverage ratio, tangible assets, operating

incomevolatility, changes of interest rate andcredit ratingbetween call date andissue date, andROA or growth rate of capitalexpenditures andR&D expenses. All

theexplanatoryvariables areas denedin the Appendix. Changein creditratingis dened asthe rating atcallminus the ratingat issue. Tocontrol for thetime and

industry effects, we also include dummy variables for each calendar year and each industry based on two-digit SIC Code. P-values are reported in parentheses.

Mean probability is the predicted probability of refunding when all explanatory variables have their mean values. Change in probability is de nes as the

percentage change in the probability of refunding when the corresponding explanatory variable is increased by one STD, and all other variables are evaluated at

their means and reported in { }.

Independent variables 1 2 3 4

Intercept 7.7941 8.1124 7.2683 7.2718

b.0001 b.0001 b.0001 b.0001

ROA1 3.8347

(0.0247)

{0.1452}

ROA2 7.5031

(0.0011)

{0.2597}

CAPEX 5.0225

(0.0317)

{0.1199}

CAPEXRD 4.9442

(0.0312)

{0.1199}

Firm size 0.4773 0.4738 0.5445 0.5445

b.0001 b.0001 b.0001 b.0001

{0.4877} {0.4898} {0.4818} {0.4817}

Tangible assets 0.0698 0.2240 0.7688 0.7751

(0.8747) (0.6065) (0.1031) (0.1004)

{0.0101} {0.0330} {0.1024} {0.1032}

Leverage (total debt) 0.0588 0.6941 0.5939 0.5833

(0.9194) (0.2690) (0.3467) (0.3549)

{0.0062} {0.0757} {0.0562} {0.0552}

Change in risk-free rate 0.0158 0.0287 0.0114 0.0119

(0.7210) (0.5212) (0.8001) (0.7930)

{0.0206} {0.0377} {0.0138} {0.0143}

Change in credit rating 0.0581 0.0805 0.0283 0.0286

(0.3239) (0.1851) (0.6375) (0.6343)

{0.0635} {0.0879} {0.0263} {0.0266}

Operating income volatility

6.8539

2.8477

7.3007

7.0060(0.4906) (0.7778) (0.5258) (0.5422)

{0.0423} {0.0178} {0.0402} {0.0386}

Mean probability 0.3355 0.3301 0.3831 0.3832

Industry dummy and calendar year dummy Yes Yes Yes Yes

Pseudo-R2 0.2949 0.3052 0.2828 0.2829

NOBS 853 833 786 786

35 Shumway (2001)shows that such a multi-period logit model (using multiple-period data before corporate events (e.g., bankruptcy or mergers) is equivalent

to a discrete-time hazard model, which produces more consistent and unbiased estimates than a static single-period logit model.36 Since we impose this lter on all our sample rms, including those that call their bonds with refund, call their bonds without refund, and not call their bonds,

it should not bias our results in any systematic way. Among 198 rms that are not present in the CRSP database at the end of 2004 and thereby excluded in our

sample, 25% led for bankruptcy by 2004.37 Since we use panel data inTable 7, standard errors in multinomial logit models are corrected for rm clustering effect.

http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-


17/20

includerm size, leverage ratio, tangible assets, operating income volatility, changes of interest rate and changes of credit rating

between call date and issue date, STD of risk-free rate, time to maturity, and measures ofrm performance and investment

activities, including ROA and growth rate of capital expenditures and R&D expenses. STD of risk-free rate is the standard deviation

of the 30-year Treasury bond yield in each calendar year. ROA or growth rate of capital expenditures and R&D expenses are the

main variables of interest.

Based onH6, we expect a positive coef

cient estimate on ROA or growth rate of capital expenditures and R&D expenses inmodels evaluating the choice of call with refund versusnot call, and a negative coefcient estimate in models evaluating the choice

of call without refund versus not call. Note that because now we are comparing the two extremes cases, call with refund and call

without refund, with the middle case, not call, we expect that our results to be not as strong as those reported in Table 6, where we

directly compare the two extreme cases.

As shown in Table 7, we nd signicantly positive coefcient estimates on both ROA1 and ROA2 in models (1) and (3),

suggesting that better performing rm is more likely to call and refund a bond rather than not call. The coefcient estimates on

ROA1 and ROA2 in models (2) and (4) are both negative but only statistically signicant in model (4), suggesting that a rmthat is

performing poorly is more likely to call a bond without a refund rather than not call. The opposite effects ofrm performance on

the choice of call with refund versus not call and call without refund versus not call is apparent, as predicted by H6. In models (5)

and (7), the growth rates of CAPEX and CAPEXRD are both positively related to the probability of call with refund versus not call,

though the relationship is only marginally signicant in model (5). In contrast, the growth rates of CAPEX and CAPEXRD are

negatively related to the probability of call without refund versus not call, and both results are signicant. This evidence suggests

that a

rm with higher investment activity is more likely to call with refund than not call, but is less likely to call without refundthan not call. This evidence again lends support to hypothesis H6.

Table 7

Multinomial logistic regressions explaining the choice to call with refund, call without refund, or not call. This table presents the coefcient estimates from

multinomial logistic regressions explaining the three choices of callable bonds: call with refund, call without refund, or not call. Our sample includes all callable

bonds during the period right after call protection expires or 1980 (whichever comes later) until the year of being called or 2004, whichever comes rst. We use

the observations ofnotcall as thebase case andevaluate theothertwo outcomes (call withrefund andcall withoutrefund) as alternatives to thischoice. Models

(1), (3), (5), and (7) evaluate the choices of call with refund and not call, and models (2), (4), (6), and (8) evaluate the choices of call without refund and not call.

Independent variables includerm size, leverage ratio, tangible assets, operating income volatility, changes of interest rate and changes of credit rating between

call date and issue date, STD of risk-free rate, time to maturity, and ROA or growth rate of capital expenditures and R&D expenses. STD of risk-free rate is the

standard deviation of the 30-year Treasury bond yield in each calendar year. All other explanatory variables are as dened in the Appendix. To control for the time

and industry effects, we also include dummy variables for each calendar year and each industry based on two-digit SIC Code. P-values are computed based onstandard errors corrected for rm clustering effect in panel data and are reported in parentheses.

Variable Call with

refund vs. not

called

Call without

refund vs. not

called

Call with

refund vs. not

called

Call without

refund vs. not

called

Call with

refund vs. not

called

Call without

refund vs. not

called

Call with

refund vs. not

called

Call without

refund vs. not

called

1 2 3 4 5 6 7 8

Intercept 0.502 3.598 1.026 3.783 0.988 3.853 1.040 3.821

(0.515) b.0001 (0.173) b.0001 (0.190) b.0001 (0.166) b.0001

ROA1 3.498 1.125

(0.016) (0.438)

ROA2 3.229 4.318

(0.074) (0.006)

CAPEX 5.977 8.735

(0.088) (0.010)

CAPEXRD 3.814 -8.118(0.230) (0.009)

Firm size 0.315 0.010 0.306 0.016 0.321 0.040 0.316 -0.042

(0.000) (0.878) (0.000) (0.799) (0.000) (0.534) (0.000) (0.522)

Tangible assets 0.369 0.212 0.490 0.342 0.681 0.043 0.649 0.034

(0.281) (0.525) (0.136) (0.295) (0.040) (0.901) (0.049) (0.920)

Leverage (total debt) -0.533 0.0 25 0.536 0.597 0.749 0.472 0.740 0.444

(0.362) (0.964) (0.376) (0.298) (0.204) (0.416) (0.209) (0.444)

Change in risk-free rate 0.000 0.071 0.000 0.066 0.007 0.066 0.004 0.066

(0.991) (0.017) (0.998) (0.029) (0.813) (0.034) (0.875) (0.033)

STD of risk-free rate 1.411 1.367 1.428 1.371 1.350 1.327 1.347 1.316

(0.003) (0.005) (0.003) (0.005) (0.004) (0.008) (0.005) (0.009)

Change in credit rating 0.018 0.051 0.017 0.060 0.029 0.025 0.025 0.025

(0.802) (0.387) (0.806) (0.320) (0.671) (0.682) (0.714) (0.692)

Operating income

volatility

12.698 14.578 14.512 16.035 17.727 19.423 17.357 19.264

(0.135) (0.067) (0.088) (0.047) (0.037) (0.019) (0.042) (0.020)

Ln (time to maturity) -1.855

1.835

1.894

1.830

1.898

1.853

1.900

1.842(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

Industry dummy and

calendar year dummy

Yes Yes Yes Yes Yes Yes Yes Yes

Pseudo-R2 0.165 0.169 0.167 0.165

NOBS 2792 2743 2704 2710



18/20

7. Conclusion

If a rm issues a non-callable bond, even when the rm's investment opportunity turns out to be poor, it may still have

incentives to invest because of the well-known risk-shifting problem. In this paper, we propose a theory that a rm could issue a

callable bond to reduce the risk-shifting problem. Because the call option enables the rm to reduce its debt obligation when its

investment opportunity turns out to be poor, this makes it more attractive for equity holders to forgo the negative NPV project and

repay the bond earlier. The cost to a rm of issuing a callable bond, however, is that it will have to incur a refunding cost if its

investment opportunity turns out to be excellent. Therefore, a rm would trade-off between the benet of reducing risk-shifting

problem and the refunding cost, when it decides whether to issue a callable versus a non-callable bond.

Our model produces several unique empirical implications that help differentiate it from others. Our empirical ndings offer

strong support to our model. We nd that a rm with poorer future investment opportunities is more likely to issue a callable

bond. In addition, a rm with a higher leverage ratio and higher investment risk is more likely to issue a callable bond. Finally, we

nd that a rm with the best performance and the highest investment activity is likely to call and refund its bond; a rm with the

worst performance and the lowest investment activity is likely to call without refunding its bond; and a mediocre rm is likely to

not call its bond. In contrast, our ndings do not seem to support the alternative theories in the literature, such as hedging interest

rates risk, signaling, solving debt overhang problems, and removing restrictive covenants.

Acknowledgments

We would like to especially thank an anonymous referee, Andres Almazan, Aydogan Alti, Ilan Guedj, Jay Hartzell, Jean Helwege,

Richard Kish,

Documents

GetWhyfirms issue callable bonds: Hedging investment uncertainty