Transcription

1 Relationship Capital and Competition in the Corporate Securities Underwriting Market Ayako Yasuda Department ofeconomics Stanford University This version: January 2001 Abstract This paper examines how relationship capital drives competition among underwriting banks in the corporate bond underwriting market. The recent deregulation and subsequent re-entry by commercial banks into this market in the U.S. raises policy interest in measuring competitive eect of the deregulation. The theories of relationship and reputation imply that (1) the entrants' products are dierentiated from those of incumbents due to their pre-existing relationships with issuers, and (2) the issuers' valuation of relationship capital in turn depends on their own reputation in the capital markets. Iinvestigate these hypotheses by using the methods developed in the empirical industrial organization literature, which have not been applied to the nancial context. In doing so the paper also attempts to ll this gap between the two literatures. Using individual issuer data, I jointly estimate a discrete choice model of underwriter demand and a model of issue-specic pricing by individual banks. I nd that there is a trade-o between relationships and price in the demand equation and that this trade-o is sharply higher for lowreputation issuers such as junk bond issuers and rst-time issuers. The lower is the reputation of the issuer, the more the demander is willing to trade o price for the relationship. This nding is consistent with the certication eect of commercial bank underwriting. Commercial bank entry has increased competition in this market because their client-specic relationship capital has increased product dierentiation in the market. And since issuers with low reputation value the relationship capital more, it is for this segment of the market that the competitive eect of the deregulation has been most dramatic. I am indebted to Tim Bresnahan, Doug Bernheim, and Roger Noll for their invaluable guidance. I also thank Masa Aoki, Serdar Dinc, Ken Judd, Manju Puri, John Shoven and all the seminar participants at the Center for Financial Studies at Frankfurt, the Federal Reserve Bank of New York, Kellogg, HBS, IMF, Stanford, UC Berkeley, UC Irvine, Wharton, and the World Bank for many helpful discussions and comments. All errors are my own. 1

2 1 Introduction This paper studies competition among banks for providing underwriting services to issuing rms in the corporate bond market. To explain the nature of this market, consider a rm wishing to issue securities. When rms issue securities, they hire a bank as an underwriter. The bank underwrites securities by guaranteeing a xed amount of proceeds that the issuing rm receives when bonds are sold to investors. In addition to assuming the market risk (and potentially the risk of reputation loss as well), the underwriter also assists the issuer with documentation of the prospectus and with marketing and selling of the bond. The issuer pays a fee to the underwriter for its service which is fully paid at the time of the issuance. Motivations for this study of the bond underwriting market in the U.S. are twofold. First, there has been an ongoing reform of banking regulations in the U.S. over the last twenty years, culminating in the recent passage of the Financial Services Modernization Act of Nearly seven decades ago, the Glass-Steagall Act of 1933 established a policy of separation between commercial banking and investment banking and, as a result, commercial banks were excluded from the corporate securities underwriting business for fty-ve years. But in recent years, there has been an eort to dismantle the separation and allow banks to engage in both businesses at the same time. As a part of this regulatory trend, commercial banks were allowed to re-enter the corporate bond underwriting business in 1987, resulting in subsequent entry by major commercial banks into the underwriting market. Given the continuing trend toward further deregulation, and the expected rise in consolidation of the industry, both regulators and antitrust authorities are interested in measuring the market power of banks in this market. To address the market power measurement issue, we need to answer a number of related questions: How are prices determined in this market? What are the sources and magnitudes of product dierentiation in this market? And how does product dierentiation found in this market aect the ability of individual banks to set prices? The price competition literature has used estimates of the demand function to quantify the ability of rms to set prices in product 1 For the literature on the banking regulation reform in the U.S., see Saunders and Walter(1994 and 1996). 2

3 dierentiated industries. 2 These questions thus motivate an estimation of a underwriter demand model, which has been delivered by neither the industrial organization nor the nance literature. This paper attempts to ll this gap in the literature. Secondly, analysis of underwriting as a bundled product with information production being one of the key cost components yields a prediction that certication ability of banks, or eectiveness in information production about the issuer, is a potential source of product dierentiation among underwriting banks. If the certication ability of banks is a signicant determinant of demand, there is a trade-o between price and certication. The entry of commercial banks into this market gives us an opportunity to empirically test this prediction. Furthermore, the demand estimates enable me to locate and quantify the market power for banks in this market. Thus the deregulation both motivates as well as provides a way of answering the question of measuring competition and market power in this market. In this paper I estimate a model of underwriter demand to investigate the following research questions: (1) Do issuing rms value the ability of underwriting banks to certify the issuer for investors? (2) If the answer to question 1 is yes, then how much are the issuers willing to trade o price for the value of the relationship? (3) Does this trade-o vary across the reputation characteristics of the issuing rms? (4) How do the abilities of banks to set prices dier between the entrant commercial banks and incumbent investment banks, and how do they vary across the reputation characteristics of the issuers? In order to estimate this model, I have constructed a data set consisting of 1,535 U.S. domestic corporate bond issues in the period This data set combines individual issuelevel and rm-level data with rm- and bank-specic data on previous loan arrangements. The underwriting fees in this market are quoted individually in each issue, and thus the fees of unchosen underwriters are unobserved and need to be imputed. Exogenous imputation based on observed prices alone will lead to selection bias which systematically underestimates imputed prices and 2 Bresnahan (1989) provides a survey of this literature. 3

4 bias the price coecient toward zero. To address this selective data unobservability issue, I jointly estimate the demand and price estimates in an iterative, EM algorithm framework. I nd that both price and pre-existing relationships are signicant determinants of underwriter choice in the bond market. Moreover, the trade-o between price and relationships varies substantially across the credit quality of issuers as well as their newness in debt capital markets. The more severe is the issuer's information asymmetry vis-a-vis investors, the more the rm values a prior relationship with a bank and the less price-sensitive it is in choosing an underwriter. In terms of economic implications on competition, own-price elasticities of underwriters are significantly lower with respect to these low-reputation rms, indicating higher mark-ups. Moreover, own-price elasticities of commercial banks are as low asthoseofinvestment banks in the junk bond category but are signicantly higher than those of the incumbents in the non-junk bond category. Pre-existing relationships with low-reputation rms are therefore a signicant source of market power for the entrant commercial banks. The remainder of the paper is organized as follows. Section 2 reviews related branches of the corporate nance literature. A model of the underwriting service market is described in Section 3. Section 4 explains the empirical specication used to estimate the demand model. Section 5 describes the data and denes the explanatory variables to be included in the estimation, and Section 6 presents the estimation results. Section 7 interprets and analyzes the economic implications of these demand and price estimates. Section 8 concludes. 2 Prior and Related Literature In the rst subsection I review related branches of the corporate nance literature. In the second subsection I discuss the previous research on underwriter competition. 4

5 2.1 Related Literature in Corporate Finance The Monitoring Role of Commercial Banks The theory of nancial intermediation has stressed the unique monitoring function of commercial banks. Anumber of papers argue that banks have scale economies and comparative costadvantages over other lenders (including individual bondholders) in producing information about the borrowers. 3 In particular, it has been argued that a bank loan renewal creates two positive externalities for the rm. It enables other fund providers to avoid duplicating the costly evaluation process, and it also provides certication of the rm's payment ability to the public. James (1987) investigates this idea by measuring the average stock price reaction of rms that publicly announce a bank loan agreement or renewal. Consistent with certication eect of bank loans, he nds that bank loan announcements are associated with positive and statistically signicant stock price reactions. Other papers, such as the one by Chemmanur and Fulghieri (1994), attribute the monitoring ability of banks to their incentive to build their own reputations as lenders. Chemmanur and Fulghieri (1994) model a rm's choice between bank loans and bonds, allowing for debt renegotiation in the event of nancial distress. The main implication of the model is that the desire of the banks to acquire a reputation for making the \right" renegotiation versus liquidation decisions gives them an endogenous incentive to devote more resources towards evaluating a rm's value than bondholders. These papers suggest that commercial banks have closer, longer-term, and more exclusive relationships with their borrowers than other types of lenders. These views support the use of pre-existing relationship variables in my model as a measure of eectiveness in information production as underwriters. In a related empirical study, Sudip Datta and Patel (1999) nd that the existence of bank debt (with any bank and not necessarily with the underwriting bank) lowers the at-issue yield for rst public straight bond oers. Several papers 4 study the implications of commercial bank underwriting using gametheoretic models. They demonstrate that even with the assumption of rational investors, a potential 3 Leland and Pyle (1977), Diamond (1984), Fama (1985). 4 Kanatas and Qi (1998), Rajan (1992a), Puri (1999). 5

6 social cost to the combining of investment banking and commercial banking cannot be ruled out. The existence and magnitude of the conict of interest problem depends on the cost of information production by the two types of banks, the timing of their access to a rm's private information, investors' beliefs about the quality of the rms underwritten by commercial banks, etc.{none of which are easily measurable. There are also a group of articles which pursue this issue empirically. These articles investigate whether there is a signicant dierence in the level of the ex post default rate and ex ante yield of bonds between commercial bank underwriters and investment bank underwriters. 5 Their statistical models yield descriptive interpretations of the data. They generally conclude that there is no detectable conict of interest in the data and in some cases there is a net certication eect A Reputation Model of the Firm's Debt Choice Diamond (1991) uses the borrowing rm's reputation in explaining the choice between bank loans and bonds. The main result of the paper is that rm reputation and bank monitoring (of the rm's investment decisions) are substitutes. The intuition for this result is as follows: Young rms and old rms without reputations tend to rely more on bank loans, because they do not have reputations to lose and therefore bank monitoring is needed to enforce ecient investment decisions. Large established rms with good reputations, on the other hand, do have a reputation to lose and therefore have sucient incentive to choose ecient investment decisions. costly, and thus this class of rms will prefer to issue bonds. Bank monitoring is The model implies that there is an intertemporal linkage between bank loans today and the 5 See Ang and Richardson (1994), Kroszner and Rajan (1994), and Puri (1994) for default rate studies using pre-glass-steagall historical data. See Puri (1996), Gande, Puri, Saunders, and Walter (1997) for studies on ex ante yield. 6 The conict of interest hypothesis (as argued historically) assumes naive investors whereas the certication eect hypothesis assumes that investors are on average rational. Since one cannot separately identify hypotheses that are based on dierent primitives of the underlying economic model, I interpret that some papers in this literature treat investors' rationality as part of the hypotheses, while others implicitly assume rational investors and therefore alter the meaning of the term \conict of interest" from that used in the original political debate. 6

7 rm's decision to issue bonds in the future: \A borrower's credit record acquired when monitored by a bank serves to predict future actions of the borrower when not monitored" (p.690). This suggests that the monitoring of rms by commercial banks in the loan market can become an asset in another market, e.g., when such banks become underwriters in the bond market. 7 Moreover, this substitution between bank monitoring and rm reputation implies that issuers with low reputation value their loan relationship with underwriting banks more than those with high reputation. 2.2 Prior Research on Underwriter Competition To the best of my knowledge, only one paper to date has studied post-entry competition issues in the corporate bond underwriting market. 8 Gande, Puri, and Saunders (1999) analyzes the eect of commercial banks' entries on underwriting fees and ex ante yield of the bonds using statistical models. Their nding of a negative correlation between the fees (and yield) and the bank entry yields a descriptive interpretation of the data. In contrast, I examine underwriter competition by estimating a choice-theoretic model of underwriter demand, treating issuers as economic agents maximizing their objective function in choosing an underwriter. This approach yields economic interpretations of the data. Specically, I measure the extent to which relationship capital is a signicant source of product dierentiation for underwriting banks. 7 Alternatively, the lending bank might bechosen as underwriter because the rm prefers to keep a single-bank relationship in all nancial services provided by commercial banks. However, the impact of the multiple- vs. singlebank relationships on credit availability, pricing, and rm performance is mixed. While Petersen and Rajan (1994) and Petersen and Rajan (1995) demonstrate that multiple-bank rms are more credit constrained (and face higher interest payments) than single-bank rms, Houston and James (1996) conclude the opposite. 8 See Hansen (1999) for recent papers on competition issues in the equity underwriting market. 7

8 3 The Underwriting Service Market When a rm decides to issue a bond, it hires an underwriting bank to market, price and distribute the security. 9 For a relatively small oering and/or if the rm is a frequent issuer, the issuer typically calls up a number of potential underwriting banks and obtain individual quotations from them. For a large issue and/or if the rm issues infrequently, itismorelikely that a \beauty contest" takes place. In such cases the issuing rm invites a certain number of potential underwriting banks to make detailed proposals in formal presentations. After being presented with these multiple proposals, the rm then decides which bank will underwrite its bond. To examine how the banks compete in this market, we need to rst understand the nature of the product that is bought and sold here. First of all, security underwriting is a bundled product. By hiring an underwriting bank, the issuing rm really buys two distinct but interrelated services: An insurance for unsold securities, and use of the bank's human resources and franchise for documentation, marketing, pricing, and sales of the security. How do banks price this mix of services? Do prices vary across transactions, and do they very across banks in a given transaction? The price variation across transactions is easily observed, and is strikingly high, especially compared to the oft-discussed \7 % x" in the stock IPO market. (See Figure 1). The variation across dierent banks in a given transaction, on the other hand, is not directly observable because the prices that unchosen banks oered are not available. But analysis of cost components in the production function of banks provides us with inference on this aspect of the competition. For the underwriting bank supplying the bundled good in a given transaction, the costs of providing these two services are positively correlated with each other and with the reputation of the issuing rm. For example, if the issuer is a \hot", well-regarded name in the market, not only is the probability of unsold securities very low, but the physical cost of marketing and selling the security is also fairly low. In contrast, it is more expensive both to insure (against unsold 9 I am abstracting away from the syndicate structure here for expositional clarity. The syndicate structure will be discussed in Section 5. 8

9 securities) and to market and distribute an obscure, less well-known issuer's bond. Investors need to be educated and persuaded harder to purchase the bond (even after controlling for higher yield of the bond), which requires that the bank's sales force also needs to be educated, etc. Thus the rst implication of the analysis is that the reputation characteristics of the issuing rms and bonds are factored into the price of underwriting services. Credit ratings of the rms and the maturities of the bonds are examples of such characteristics. Second, productions of these two services are complementary, because they share a common cost component, that is, the cost of assessing the issuer's creditworthiness and certifying the ndings to the investors. I refer to this assessment and certication aspect of the underwriting service as information production. Measuring how eective banks are in producing information about the issuer is fundamental to the analysis of competition in this market. When certain banks are perceived by investors as more eective in information production, they can build up demand faster for the securities they sell and thus will face lower risk of unsold securities and lower cost of marketing and sales as well. Thus heterogeneity among banks in information production eectiveness implies that they face dierent cost functions even after controlling for the issuers' reputation characteristics, which leads to the second implication that prices vary across banks for a given rm as well. From the point of view of issuing rms, which are the demanders in this market, banks with superior information production eectiveness may be preferred for two reasons. First, since unsold securities or otherwise failed transactions are likely to hurt their reputation in the capital markets for future transactions, they value certication ability of underwriting banks to lower the risk. Second, the superior eectiveness of banks in building up demand for the security may indicate that they can negotiate a lower yield for the bond than other underwriting banks. So the rst prediction about the competition in this market is that the product in this market is dierentiated mainly along two dimensions price, and information production eectiveness. This is the analytical basis of the demand model used in this paper. One way in which banks come to possess more eective information production technology is by having built relationships with the rms. Having already networked and established commu- 9

10 nication channels with an issuer increases a bank's eectiveness in producing information about that particular issuer. As reviewed in Section 2, the nancial intermediation literature provides ample theory and evidence that loan relationships are exclusive and long-term in nature, thus making commercial banks an \insider" with respect to the rm. 10 The recent re-entry of commercial banks into this market gives us an opportunity to test the prediction on the demand behavior of the rm, because these entrant banks have had pre-existing relationships with some of the issuers built through their loan business prior to the deregulation. As with prices, there are variations in these relationships both across issuers and across banks. I treat these relationships as predetermined and exogenous to the competition in the underwriting market, and use the variation in price and relationship to estimate a model of underwriting demand. The price competition literature has used estimates of the demand function to quantify the ability of sellers (banks in this paper) to set prices in product dierentiated industries. The rst-order condition of a bank's prot maximization problem gives the basis for using own-price elasticities to estimate the mark-ups that sellers charge. If there is a high trade-o between relationships and price in the demand equation, that implies low price elasticities for rms with relationships and thus high market power. Lastly, the analysis of price determination and demand behavior yields another prediction that this trade-o between price and relationship varies with reputation of the rm. Firms with low reputation stand to have the most gain from choosing an underwriting bank whose eective information production capability can certify them. For rms with high reputation, on the other hand, the information production capability of banks is largely redundant, since their security can sell easily in the market regardless of who the underwriter is. In other words, it is predicted that there is a higher trade-o between relationship and price for rms with lower reputation. If true, this indicates that not only are the pre-existing relationships a source of market power for the entrant commercial banks, but their signicance is inversely related to the reputation of issuing rms. 10 Leland and Pyle (1977), Diamond (1984), Fama (1985), Rajan (1992b). 10

11 4 Empirical Specication of the Demand Model At the heart of my analysis lies a discrete choice model of a rm's underwriter choice. The approach taken here is transaction-based, rather than capital stock-based. This follows the pecking-order nancing model 11 and is consistent with the premises of other debt choice models as well. 4.1 Notation Let the banks be numbered 1::K, indexed by k. Let the rms be numbered 1::N, indexed by i. 4.2 Model Based on the analysis of price determination in this market, the price (observed or not) oered by bank j to rm i is specied as p i j = z T i j 1 + z i j T j 2 + i j (1) where i j N(0 2 )iid.(for simplicity of notation, z T k is used to refer to the non-random part of the price from now on.) z i here capture the reputation/risk characteristics of issuers/bonds which determine the cost of underwriting for the banks as discussed in the previous section. Examples of such characteristics are credit ratings and maturities of bonds. The subscripts in j reects the implication of the analysis that individual banks price these characteristics of issues/bonds dierently. The price equations are thus specied in order to impute the unobserved prices. Imputation methods are discussed in Section In this well-known partial equilibrium model of the rm's nancing behavior, the rm optimizes with respect to each nancial transaction, taking its need for external nance as given. The model implies a hierarchical decision tree for the rm, which is easily adaptable in a nested-logit model framework. See MacKie-Mason et al. (1990) Helwege and Liang (1996) for use of nested-logit models in testing pecking-order model predictions. 11

12 Consider a rm i that chooses an underwriter after observing p i k for k = 1::K. The product in this market is dierentiated along two dimensions, namely, price and certication ability, which is measured by pre-existing relationships in the loan market. I specify a nested-logit model according to the nest structure given in Figure 2. The lower nest refers to the choice of a particular underwriting bank, whereas the upper nest refers to the choice between commercial banks and investment banks (indexed by m). The indirect value function for rm i choosing bank j which belongs to upper nest m is given by: V i (m j) = X r Y i r [d T j + r p i j + r x i j ]+ w T i m + i (m j) p i j is the underwriting fee charged by the bank x i j is a prior loan relationship variable d j isabankdummyvector Y i r is an indicator variable for the reputation characteristics of the issuer w i are the bond/issuer characteristics i (m j) is the idiosyncratic rm-bank error The multiplication of [d T j + rp i j + r x i j ] by Y i r and summing over r indicates that coecients and are allowed to vary across the reputation characteristics, which is indexed by r. I assume that all issuers face the xed choice set k = 1::K. The actual number of underwriting banks in the data sample is only 16, so to assume that issuers know all of them is not unreasonable. I also assume that the constant term and relationship variables suciently pick up most of the unobservable quality characteristics of banks so that any unobservable quality terms left in the residual terms are small and can be ignored. The relationships are client-specic qualities of the banks, whereas the constant term captures any xed quality of banks that all issuing rms perceive identically. Assume that the error term follows the Generalized Extreme-Value (GEV) distribution. 12

13 McFadden(1978 and 1981) showed that the assumption of the GEV distribution implies The lower-nest choice probability: Pr(jjm Y i r )= e dt j + rp i j + r x i j P Kk=1 e dt k + rp i k + r x i k (2) The upper-nest choice probability: Pr(m) = ewt i m + I i m Pt ewt i t + I X i t where I i t = log( e dt l + p i l + x i l) l2l t The inclusive value I i t measures the expected aggregate value of subset t and the coecient reects the dissimilarity of alternatives within a specic subset. To be consistent with value maximization, must lie within the unit interval. A value of 1 implies that the nest is redundant, whereas value 0 implies that the error terms within the given nest are perfectly correlated. One implication of multinomial logit models is that the odds ratios between choices exhibit the so-called IIA property. It means that the ratio P j =P k for any given choices j and k is independent of the remaining probabilities. Nested-logit models partially relax this restrictive property by allowing exible substitution patterns between choices across nests. One of the parameters of interest in the model is the nesting parameter. Within the context of the nested-logit model, captures the degree to which error terms within the nest are more highly correlated than error terms across the nests. As mentioned in the previous section, there is a logical link through this nesting parameter between the two hypotheses examined in the related empirical literature the \Certication Eect" and \Conict of Interest" views and my model here. There are two separate theories that explain how underwriting services performed by commercial banks may be dierent from underwriting performed by investment banks. If the 13

14 issuers' valuations of services dier systematically between commercial banks and investment banks, then the presence of either one or both of these eects posited in the theories would cause to tend toward zero. If, on the other hand, implications of neither of theses theories are signicant enough, then we will fail to reject the hypothesis that is Research Hypotheses Revisited The empirical specication of the demand model corresponds to my research questions as follows: (1) Do issuing rms value the ability of underwriting banks to certify the issuers for investors? This is captured by coecient in Equation 2. (2) How much are the issuers willing to trade o price for the value of the relationship? This is measured by the ratio of the coecient. (3) Does this trade-o vary across the reputation characteristics of the issuing rms? This hypothesis is tested by interacting the demand function with the reputation characteristics of the issuers, thus allowing the valuations of relationships and price to vary across these characteristics. (4) How do the abilities of banks to set prices dier between the entrant commercial banks and incumbent investment banks, and how do they vary across the reputation characteristics of the issuers? This is measured by calculating own-price elasticities of individual banks from the demand estimates. 4.4 The Estimation Method A data issue arises in studying this market because prices vary across both issuers and banks, but only one price per issue, that is, the price oered by the bank which is hired to underwrite the bond, is observed. So not only are there missing price data, but there is also a selection-driven endogeneity in the observed price data. Let c i represent the index of the bank chosen by rm i. Since price enters the value function of the issuing rm negatively, the fact that a given bank 14

15 was chosen over other banks in the choice set implies that these observed prices, (p i j j = c i ), are on average lower than the unconditional distribution of p i j. As a result, if we impute unobserved prices by obtaining estimates of from Equation 1 using observed prices as dependent variables, we will be systematically underestimating unobserved prices and biasing the price coecient toward zero, or even positive. To address this selection issue, I estimate the parameters of the lower nest of the underwriter choice model using an EM algorithm framework 12. The main idea is to estimate price estimates and demand estimates jointly in an iterative algorithm where price imputation is done conditional on the information on c i i =1::N. The main appeal of using this framework for my data problem is that it provides an iterative procedure where, if not for the systematic absence of some data, the Maximum Likelihood estimation is straightforward. The demand estimates obtained from this estimation method are then used to estimate the upper level of the nested-logit model. Details of this method are found in Appendix. 5 The Data 5.1 Data Sources I construct the dataset using two data sources. One is U.S. Domestic New Issues Database by Thomson Financial Securities Data which compiles new issues information from company lings, press releases, and news sources. The other is the Loanware Database compiled by Euromoney, which is a global database on syndicated loan markets that contains U.S. domestic syndicated loan transactions as well as non-syndicated, single-bank transactions. A full list of variables used in the estimation is provided in the Appendix. 12 Dempster, Laird, and Rubin (1977), McLachlan and Krishnan (1997). 15

16 5.2 Data Selection I choose the sample period to be from January 1, 1993 to August 31, 1997 for roughly 4 2=3 years. The criteria I used are as follows. First, the sample should begin after January 1989, when the rst commercial bank underwriting of a public corporate bond took place. Second, the economic and regulatory environment surrounding the underwriters and issuers should also remain relatively stable. The decision to omit data after August 1997 is primarily due to the merger in September 1997 of Salomon Brothers and Smith Barney, as a result of the acquisition of Salmon by Travelers, Smith Barney's parent company. Both Salomon and Smith Barney were leading investment banks in the U.S. corporate bond underwriting market at the time of the merger. It is of course quite infeasible to omit all mergers and acquisitions activities involving underwriters from any reasonably long sample period. In the wake of nancial globalization, there have been a urry of cross-border and cross-industry mergers between major players. A few examples include the Paine Webber Group's purchase of Kidder, Peabody & Co. in Oct. 1994, and Swiss Bank Corp.'s double acquisitions of the investment-banking unit of S.G. Warburg in May 1995 and of Dillon, Read &Co. in May However, none of these earlier M&A events involved two rms ranked in the top ten of the corporate bond underwriting market. Rather, they are characterized as acquisitions of veteran Wall Street players by commercial banks just entering the underwriting market, so one can argue that ownership changes of these rms had no substantial impact on the competitive environment in the underwriting market. In contrast, the Salomon-Smith Barney merger represented the rst merger of two existing major investment banks, thus aecting the overall market structure. I exclude nancial rms (one-digit SIC code 6) and regulated industries (one-digit SIC code 4) from my study. I also concentrate on the top sixteen underwriters 13 of xed-rate, non-convertible corporate debt. Fixed-rate debt comprises about 90% of all non-convertible issues. Moreover, I nd that both the composition and the sum of market shares of top underwriters are virtually uniform between xed-rate and other coupon-type bonds. For my sample, ve of the sixteen underwriters are Section 20 subsidiaries 14 of bank holding companies, namely J.P. Morgan, Chase Manhattan 13 The rankings were based on the dollar value of underwritings and gave full credit to the book-runner(s). 14 The name is derived from Section 20 of the Glass-Steagall Act, which prohibits member banks in the Federal 16

17 Bank, Bankers Trust, Citicorp, and Nations Bank. Using the above criteria results in a sample of 1,535 non-convertible, xed-rate corporate bond issues. 5.3 Variables in the Demand Equation The Price Variable The price variable (PRICE) used in the estimation is a gross spread expressed as a percentage of the size of the bond. The gross spread is the fee that underwriters receive, or the dierence between the price at which securities are sold to investors and the price paid by the underwriters to the issuing rm. Atypical public bond oering consists of multiple underwriters forming a selling syndicate. As shown in Figure 2, one underwriter serves as the book-runner (or the lead-manager) who organizes and managers the deal. I identied the book-runner (or the lead-manager) as the underwriter of a given issue. 15 Given the syndicate and fee structure as illustrated below and shown in Figure 2, this seems to be a reasonable assumption. This rule is also consistent with the perception of practitioners who advertise their bank's market position in terms of \book-runner ranking" The Loan Variable I construct dummy variables LOAN1-LOAN16 using transactions data from the Loanware database. These variables capture the presence of pre-existing loan relationships between a given rm and individual commercial banks that were established prior to the entry of the banks into the underwriting market. (see Appendix A for the exact variable denitions). I treat these relationships as predetermined and exogenous to the competition in the underwriting market. Reserve System to be aliated with rms which are engaged \principally" in the securities business. These subsidiaries were permitted to enter the investment banking business by meeting various requirements, such as a ceiling on their security-related revenues. 15 In a small number of cases where there were two co-book-runners, each was counted separately as if they underwrote separate issues. 17

18 A loan agreement frequently (but not always) consists of participation by a number of banks in a way that is analogous to the structure of a selling syndicate in a public bond oering. Thus I use arrangership of loan agreements rather than capital participation in a loan as an indication of a \prior banking relationship". On rare occasions, some top-tier investment banks also act as arrangers of syndicated loans. This variable needs to be interpreted dierently from an arrangership of syndicated loans by commercial banks, because investment banks do not participate in the syndicate as creditors, whereas a commercial bank arranger is usually the top lender in the syndicate as well (this is another justication for using an arrangership as a measure of a relationship). My interpretation is that this variable for investment banks is potentially an indirect measure of their closeness to the issuer through their investment banking activities, of which loan syndication is a relatively small but recently growing part. But it is also plausible that they are being chosen as arrangers only because they know other banks well instead of being committed to knowing the borrowing rm. If the latter factor dominates the former, the LOAN variable for investment banks should not be signicant in the demand model Variables in the Price Equation As discussed in Section 1, underwriting fees are determined in part by various costs, including distribution costs, the expected cost of taking market and reputation risks, and information production costs. I estimate my model using various specications for the price equations. The basic variables that I include in all specications are a constant term, CREDIT (a dummy variable for a junk-bond category), and some maturity variables that are non-linear in the year of maturity. Being in the junk-bond category means issuers have less nancial strength and in general are lower reputation than those in the non-junk bond category. This increases the risk-related cost for the underwriter. It might also mean that it is more costly to distribute these bonds because the company is less well-known and investors need to be marketed more intensively (which also feeds back to create 16 I informally explore this question in Section 6 after the main results. However, this variable is most likely to be endogenous with respect to the choice of investment banks as underwriter, since for investment banks bond underwriting is their main business, whereas loan arrangership is a new business area that they have entered only in the last several years. 18

19 potentially greater market risk). For similar reasons, investors require substantially higher yields for junk bonds. In general, underwriters also demand higher underwriting fees for longer maturity bonds. This makes sense to the extent that a normal yield curve is also upward sloping in addition, the secondary market for 30-year corporate bonds is much less liquid than for 30-year treasury bonds. This is measured by including some non-linear functions (dummy variables, log, etc.) of maturity in years. In some specications, I also include a variable that represents the previous issue experience of rms in the bond market. From the underwriter's point of view, frequent issuers are easier to market than rst-time or infrequent issuers. They are also less likely to lead to failed transactions because the track record of previous issues gives a reputation to the name of the issuer. Thus the price is expected to be decreasing in this variable. Another indicator of frequency of issues is whether the bond was issued under shelf registration or the Medium-Term Notes (MTN) program. Registering for this program simplies the ling process and reduces the legal and accounting costs of incremental issues. I used the sample of observed (and thus chosen) prices to check whether these variables aect the price in the predicted direction, and in fact they do Descriptive Statistics of the Sample The tables presented here give descriptive statistics for the sample used in the estimation. In all of these tables, the rst two columns segment the data according to whether the underwriter was a commercial bank or an investment bank. In Table 1 the sample is tabulated by issue size, maturity, and credit rating. Several observations can be made. First, bank-underwritten issues are relatively small compared to investment bank-underwritten issues. Their maturity also tends to be slightly shorter, but no better or worse in terms of credit ratings. There are a few plausible reasons for this. For example, if a smaller, 17 I used these OLS estimates as starting values for price equation parameters in the estimation. Coecients have the predicted signs and are signicant. 19

20 younger rm is more likely to choose the commercial bank with which it had built close ties, the issue size might beproxying for characteristics of that rm. Or if commercial banks have a smaller distribution capability relative to investment banks, the issue size might then be reecting the supply-side constraint. In Table 2, the sample is tabulated rst by previous issue experience and then by the issuer's SIC code. Commercial bank issues are relatively more frequent among rst-time issuers. This observation is consistent with the rst explanation given above for the size dierence. In contrast, there is little dierence between bank and investment bank subsamples in terms of the distribution of issuers across dierent industries (Table 3). 6 Estimation Results The estimation results of the base model are presentedintable The base model is where there are no interactions between valuations of price and relationship variables with issuer's reputation characteristics. The price coecient is negative and signicant, whereas the relationship coecient is positive and signicant. Thus both price and prior loan relationships are signicant determinants of demand in the underwriting market., the dissimilarity coecient of the nested-logit model, is , which is not signicantly dierent from one at the 5% signicant level by the Walt test. It is signicantly dierent from zero. = 1 statistically implies that the nesting as specied in Figure 2 is redundant. Besides the inclusive value variables, I also include issuer and bond characteristics in the upper nest. Since these are chooser-specic variables, parameters are estimated separately for each choice. However, the coecients for one choice (in this case investment banks) are normalized to zero, so the reported coecients are for the choice of commercial banks. 18 For this and all other model specications, the nested-logit demand model was estimated using a sequential estimation method consisting of two successive logit steps. The standard errors in the second stage were calculated using the formula given in McFadden (1981). 20

21 The estimates of equations determining prices,, are reported for each bank. Commercial banks and investment banks are listed separately, and within each category the banks are listed in the order of their market shares in the sample. Bank1 is therefore the top-ranking commercial bank underwriter, whereas Bank6 is the top-ranking investment bank underwriter and is the topranking bank overall. Coecients for the maturity and junk bond variables are mostly positive, whereas those for the past issue experience and shelf-registration variables are mostly negative. These ndings are consistent with the analysis of price determination in Section 3 and with the discussions of variables entering pricing equations in Section 5. Of these four reputation characteristics of issuers/bonds, the coecients for junk bond variables are much larger in absolute values than others. JUNK BOND and MTN are indicator variables, and the sample mean of variables for maturity and past issue experience are and 1.118, respectively. Thus the eect of being in the junk bond category on price dominates the other categories. This coecient also tends to be smaller for top-ranking banks. Some of the coecients also vary signicantly across banks within a given category. For example, Bank1 and Bank2 have a dierential of almost 0.8 for the junk bond variable. As JUNK BOND is an indicator variable, this translates to a price dierential of expressed in percentage of the issue amount, or, evaluated at the mean issue size of $180 million, approximately $1.44 million. The dierence between Bank3 and Bank8 for the maturity variable is similarly large, at The sample mean value for the maturity variable is 2.338, so this translates to about $1.36 million. The coecients for pre-existing loan relationship variables are mixed in sign and vary widely in magnitudes as well. Table 5 reports the estimation results of one of the interacted models where price and relationship coecients are allowed to vary across the reputation characteristics of issuers, namely, credit rating. The criterion I use is whether the issuer's credit rating is in the junk-bond category at the time of the issuance. Being in the junk-bond category means issuers have less nancial strength and in general lower reputation than those in the non-junk bond category. The price coecient 1 for non-junk (i.e. \high quality") issuers is sharply more negative at , whereas the price coecient for junk-bond issuers (\low quality") is negative but a lot smaller in magnitude at

22 The loan coecients r are both positive as predicted, though essentially equal in size. The dissimilarity coecient is again not signicantly dierent from one. 1 is positive, indicating that rst-time issuers are more likely to choose a commercial bank than seasoned issuers. 19 2isnegative. If the issue amount isproxying for the size of the rm, this suggests that smaller rms are more likely to choose a commercial bank underwriter than larger rms. The price estimates presented in Table 5 are qualitatively similar to those in the base model in Table 4. The quantitative changes in estimates from the base model to interacted model, however, are concentrated in the commercial bank category and Bank11. Bank11 is not a highranking bank overall, but is the top-ranking underwriting bank in the junk bond category. Thus we can see that allowing for variation in the demanders' trade-o between relationship and price in the model specications aect the price estimates of certain banks more than others. In Table 6, I report the estimation results of another interaction model, where the tradeo between price and relationship in the demand equation is allowed to vary along newness of the issuers in the corporate bond market. Investors are less likely to be familiar with or even recognize the name of rst-time issuers in the market, so these rms are worse o than seasoned issuers in terms of their market reputation. Seasoned issuers, on the other hand, have a track record of issuing public debt, which contributes positively to their reputation. The price coecient 1 for non-rst-time (i.e. \high reputation") issuers is negative at , whereas the price coecient 2 for rst-time issuers (\low reputation") is negative at , less than half in size. The loan coecients 1 and 2 are both positive as predicted, with 2 being about 50% larger. Both results are consistent with the prediction of the model. The size of the dissimilarity coecient is close to 1, as in the two previous specications. 19 I also ran a dierent specication of the upper nest where there are two separate nests for banks with and without pre-existing loan relationships. While the coecient for FST TIME for banks with pre-existing relationships is signicantly positive and even larger than the coecient in the basic model, the coecient for FST TIME for banks without relationships is negative. This indicates that the rst-time issuers are more likely to choose commercial banks over investment banks despite the formers' recent entry into the market, but only if they have had pre-existing relationships with these banks. 22

23 Price estimates are also qualitatively similar to the other specications. 7 The Economic Implications of Estimation Results The demand estimates presented in Table 4{6 are consistent with the predictions of the model. Both relationships and price are signicant determinants of underwriter demand in the market. The trade-o between relationships and price indicates that issuers are willing to pay higher price for underwriting services from banks with pre-existing relationships. The trade-o is quantied by the ratio of the two coecients,. In the base model reported in Table 4, this ratio is Since price is expressed as a percentage and the relationship is an indicator variable, this ratio has a unit of 0.805%. In terms of the underlying demand model, a bank with a relationship can charge a premium of up to 0.805% before an issuer prefers a bank without a relationship (holding all else constant). Evaluated at the sample mean issue size of about $180 million, this translates to about $1.45 million in dollar value. Since the level of the underwriting fee paid by the issuers in the sample ranges anywhere from $200,000 to several million dollars, the value of a relationships implied by the results is quite substantial, and at the same time reasonable. In Table 5 where this trade-o is allowed to dier between non-junk bond and junk bond issuers, the ratios are and for high- and low-quality issuers, respectively. Evaluated again at the mean issue size of $180 million, the valuation of pre-existing relationships for two classes of issuers are approximately $900,000 and $5.5 million, respectively. The 0.501% for high-quality issuers roughly matches the mean of the underwriting fee in that category, whereas 3.051% is on the high end of the range of underwriting fees paid by junk bond issuers. The large dierence in the values of r r conrms the prediction that there is an inverse relationship between the reputation of issuing rms and their valuation of the ability of underwriting banks to certify them for investors. Similar implications are obtained from results reported in Table 6, with the ratios being and for non-rst-time and rst-time issuers, respectively. Evaluated at the mean issue 23

24 size of $180 million, the valuation of pre-existing relationships for these two classes of issuers are approximately $900,000 and $1.8 million, respectively. The higher trade-os for low-reputation rms suggest that banks derive greater market power from these pre-existing relationships when providing services to low-reputation rms. To quantify this economic implication, I calculated the own-price elasticities of individual banks from the demand estimates. Table 7 reports the results of calculations based on the estimates reported in Table In the rst column I report the mean own-price elasticities of sixteen individual banks where all observations were used for calculations. On average the entrant commercial banks face a higher price elasticity of compared to the investment bank average of , indicating less market power than the incumbents. When calculations are done separately for non-junk and junk bond issuers, however, a more subtle picture of competition emerges. First, the price elasticities that banks face when supplying underwriting services to junk bond issuers are sharply lower, roughly half of the elasticities for non-junk bond issuers. Second, comparing the means of two peer groups, I nd that the price elasticities that commercial banks face with respect to junk bond issuers are as low as those of investment banks ( and , respectively), whereas the entrants face higher price elasticities relative to incumbents, compared to , when facing non-junk bond issuers. Therefore, the pre-existing relationships with low-reputation rms are a signicant source of market power for the entrant commercial banks. Despite their recent entry into the market, the commercial banks are on a par with the incumbent investment banks when selling their services to low-reputation issuers. In Table 8, I calculate a percentage change in own total demand from 1% increase in price across the sample. The results and implications are qualitatively quite similar to Table 7. It is also notable that there is a signicant variety in the banks' ability tosettheir price within each group. There seem to be two tiers within the investment bank group, with Bank6 to Bank10 constituting the rst-tier with greater market power, with Bank12 to Bank16 lagging behind. The only exception is Bank 11, which does well in the junk bond category but not in 20 Since was not signicantly dierent from one, the elasticities were calculated with only the lower-nest results. So these calculations are essentially based on a logit model specication. 24

25 the non-junk bond category. In the commercial bank group, only Bank1 seems to belong to the \rst-tier" class. Bank3 and Bank5 seem to t the analytical prediction of the entrant commercial banks the best, enjoying comparably greater market power with respect to junk bond issuers but not with respect to non-junk bond issuers. Bank2 and Bank4, on the other hand, seem to lag behind overall. In fact, once I control for rm-and-bank-specic certication ability of commercial banks, I nd that there does not seem to be any organizational form-specic \quality" left that makes the entrant commercial banks dierent from the incumbent investment banks. This is reected in the size of, the dissimilarity coecient. Commercial banks as a group per se are neither closer substitutes for each other nor more imperfect substitutes for investment banks. Instead, what matters is the client-specic relationship capital that makes them locally more eective in producing information, and the valuation of such relationships by issuers, which in turn depends on the reputation of the issuers in the capital market. Finally, the price estimates reported in Table 4-6 are open to some economic interpretations. First, it is interesting to note that coecients for JUNK BOND are all positive but tend to be smaller in size for top-ranking banks. This might be because the top-ranking underwriting banks are also the largest banks with the best and largest sta, and that enables them to be much more eective in their production, leading to lower cost. It is possible that their cost savings are suciently large that their junk bond premium in fee is lower even with the higher mark-ups implied by their lower price elasticities. Another possibility is that the assumption of a xed 16-bank choice set in the demand model is highly restrictive for this class of issuers. In other words, the variation in the JUNK BOND coecients are consistent with a market structure where the rms that I observe to be underwritten by low-ranking underwriters did not have top-ranking underwriters in their choice sets. This is also consistent with the the ndings that all banks, including even the low-ranking banks, exercise signicant market power in the junk bond category. While it is conceptually appealing to relax the assumption of a xed choice set across all issuers, it is not clear whether there is any publicly available data that enables researchers to make inference about issuer-specic choice sets One possibility that I am aware of and planning to explore is to use the information on syndicate structure of 25

26 8 Conclusion This paper has examined competition in bond underwriting markets by measuring the ability of individual banks to set prices in the presence of buyer-and-seller-specic relationships. Underwriting is a bundled product where underwriters insure the issuing rms against unsold securities as well as provide their franchise capacity in marketing, pricing and selling the security. Both aspects or the product requires underwriting banks to produce information about the issuing rm, that is, assess and certify the rm for the investors. Issuing rms, demanders in this market, value eectiveness of underwriting banks in information production to the extent that such ability of banks can add to the success of their oerings both in terms of reduced risk of unsold securities as well as lowered yield. This implies that there is a trade-o between information eectiveness (a quality measure) and price in the demand equation. Recent re-entry by commercial banks into this market raises policy interest in measuring market power of banks in this industry as well as provides an opportunity to test the predictions of the model. Pre-existing relationships with rms through their loan business make the entrant commercial banks potentially more eective information producers for those specic issuers. Based on this analysis of the underwriting service market, I estimate a model of underwriter choice using rm-level issue data of corporate bonds in the U.S. domestic market. The estimates allow metoquantify the trade-o between relationships and price that demanders make, and I use the demand estimates to measure variation in the ability of individual banks to set their prices. A discrete choice model of underwriter demand and a model of issue-specic pricing by individual banks are jointly estimated. I nd that both price and pre-existing relationships are important determinants of underwriter choice. The trade-o between relationships and price are considerably high relative to the level of fees being paid in this market. Moreover, this trade-o varies substantially and systematically across the reputation characteristics of the issuers. The lower is the reputation of the issuer, the more the demander is willing to trade o price for the relationship. This nding is consistent with the certication and information production model the issues. For example, I may try a specication where only those banks in the nal syndicate are allowed to be in the choice set of the issuer. Names of co-managers are available in the Thomson Financial New Issues Database. 26

27 of underwriting. Issuers with low reputation, such as junk bond issuers and rst-time issuers in debt capital markets, stand to gain themostfromchoosing an underwriting bank whose eective information production capability can better certify them and therefore are willing to pay for it. For rms with high reputation, on the other hand, the information production capability of banks is largely superuous, since their security can sell easily in the market independent of who the underwriter is. The demand estimates are used to calculate the ability of individual banks to set prices. I nd that own-price elasticities of underwriters are signicantly lower with respect to low-reputation rms, indicating higher mark-ups. Moreover, own-price elasticities of commercial banks are as low as those of investment banks in the junk bond category but are signicantly higher than those of the incumbents in the non-junk bond category. Despite their recent entry into the market, the commercial banks are on a par with the incumbent investment banks when selling their services to low-reputation issuers. Pre-existing relationships with low-reputation rms are therefore a signicant source of market power for the entrant commercial banks. Finally, once I control for client-specic certication ability of commercial banks, I nd that there does not seem to be any organizational form-specic \quality" left that makes the entrant commercial banks dierent fromtheincumbent investment banks. Commercial banks per se are neither closer substitutes for each other nor more imperfect substitutes for investment banks. Instead, what matters in the post-entry competition is the client-specic relationship capital that makes them locally more eective in producing information, and the valuation of such relationships by issuers, which in turn depends on the reputation of the issuers in the capital market. A List of Variables (A) Dependent Variables BOOK: Takes a value of i if the issuing rm chooses bank i in the given observation. ( i = 1-16) 27

28 BANK DUM: Takes a value of 1 if the chosen bank in the given issue is a commercial bank 0 otherwise. (B) Choice Characteristics: LOAN1-LOAN16: LOAN i Takes a value of 1 if bank i ever acted as an arranger in a loan agreement for the rm in the given issue in the period prior to the sample period ( ) 0 otherwise. PRICE1-PRICE16: PRICE i is the gross spread (expressed as a percentage of the issue amount) charged by bank i for the given issue. (C) Issuer/Bond Characteristics: FIRST TIME: Takes a value of 1 if the issuing rm in the given observation never issued a non-convertible bond in the domestic U.S. bond market between otherwise. AMOUNT: The size of the issue in 100-millions of U.S. dollars. LMAT: natural log(maturity of bonds in number of years). JUNK BOND: Takes a value of 1 if the given issue's credit rating is below Baa based on Moody's credit rating 0 otherwise. LNBOND: natural log(number of previous bond issues + 1). MTN: = 1 if the issue is under the Medium-Term Notes (MTN) program, or else 0. (D) Other: INC0-INC1: INC1 is the inclusivevalue for Commercial Banks at the C-Bank/I-Bank node of the nest. INC0 is the inclusivevalue for Investment Banks at the C-Bank/I-Bank node. 28

29 B Estimation Method In the EM framework, the observed data are viewed as being \incomplete", and are augmented by unobserved data to make up the \complete data". Each EM iteration involves an E-step where the conditional expectation of the complete-data log likelihood given the observed data is computed using the previous estimates (0), and an M-step where the conditional expectation is maximized over. This procedure is repeated in an iterative manner until convergence is achieved. Let c i represent the index of the bank chosen by rmi. Let = f g. We observe c i and p i ci, as well as x i j, z i. The task is to estimate according to the maximum likelihood principle. We will do this by an EM-type algorithm, assuming p ;ci to be the \hidden" data and hence fc i p ci p ;ci g to be the complete data. Thus we need to establish Pr(c i p ci p ;ci j). Pr(c p c p ;c j) =Pr(cjp c p ;c )Pr(p c p ;c j) by Bayes rule. According to the logistic choice model Pr(cjp c p ;c )= e dt c +pc+xc P Kk=1 e dt k +p k+x k According to the iid normal distribution of k,weknowthateach p k N(z T k 2 ) independently. Hence Pr(p c p ;c j) = KY k=1 1 p 2 2 e; (p k;z T k ) 2 Hence, we have log likelihood of the complete data (of a single rm) as ln Pr(c p c p ;c j) = ; ln KX k=1 (p k ; z T k ) 2 ; K 2 ln 22 + e dt c +pc+xc P Kk=1 e dt k +p k+x k (3) 29

30 In order to implement the E-step, we need to compute E (0)(lnPr(c p c p ;c j) j c p c ) = = Z 0 Z Z ln(pr(c p c p ;c j)) Pr(p ;c jc p c (0) ) dp ;c Q k6=c e; (0) (p k;z T (0) k )2 e (0) p c+ (0) x c + P ln(pr(c p c p ;c j)) k6=c e(0) p k + (0) x k dp ;c 0 1 C A ;1 Qk6=c e; (0) (p k;z T (0) k )2 1 C e (0) p c+ (0) x c + Pk6=c e(0) p k + (0) x k A dp ;c Note that the rst integral term is irrelevant in the M-step because it is a function only of the old parameters (0) and therefore is invariant with respect to new. So for the rest of this section I will drop this term from the analysis. What remains inside the second integral term is the product of a log of complete-data likelihood (evaluated at the new ) and the remaining part of the conditional probability Pr(p ;c jc p c (0) ), to be evaluated at the old. These are high-dimensional (K = 15)integrals over hybrid distributions consisting of normal and logit components and are computationally non-trivial. Neither numerical integration nor Monte-Carlo EM (where the E-step is replaced by a Monte-Carlo process) is trivial nor immediately promising given the high dimensionality. So instead I use what is commonly referred to as an \EMtype algorithm," whereby the single most likely value p ;c that maximizes the conditional density above (i.e. only Pr(c p c p ;c j (0) )) is computed and a probability of1 is placed on this data. In terms of the underlying economic problem, this part can be described as adjusted price imputation, where instead of using unconditionally imputed prices for unobserved prices, I am replacing them with prices that are adjusted so as to maximize the joint likelihood Pr(c i p c p ;c ), using estimates of from the previous iteration. To monitor convergence, we needtoevaluate the observed-data likelihood function L( (k) ) in each (kth) iteration. In my model the incomplete-data likelihood function is expressed as Pr(c p c j) = = Z Pr(c p c p ;c j) dp ;c Z KY k=1 1 p 2 2 e; (p k;z T k ) 2! P epc+xc Kk=1 dp ;c e p k+x k 30

31 As discussed in the previous section, the integrals above are computationally challenging. Laplace's method provides a useful way of approximating integrals that take the form I() = Z D e ;g(x) f(x) dx where is a large parameter 22. I apply this approximation method to evaluate the observed-data likelihood function. References Ang, J. S. and T. Richardson (1994). The underwriting experience of commercial bank aliates prior to the Glass-Steagall Act: A re-examination of evidence for passage of the Act. Journal of Banking and Finance 18 (2), 351{95. Bresnahan, T. (1989). Empirical studies of industries with market power. InR. Schmalensee and R. Willig (Eds.), Handbook of Industrial Organization, vol. II. Amsterdam: Elsevier. Chemmanur, T. J. and P. Fulghieri (1994). Reputation, renegotiation, and the choice between bank loans and publicly traded debt. Review of Financial Studies 7 (3), 475{506. Dempster, A. P., N. M. Laird, and D. B. Rubin (1977). Maximum likelihood from incomplete data via the EM algorithm (with discussion). Journal of the Royal Statistical Society B 39, 1{38. Diamond, D. (1984). Financial intermediation and delegated monitoring. Review of Economic Studies 51, 393{414. Diamond, D. W. (1991, August). Monitoring and reputation: The choice between bank loans and directly placed debt. Journal of Political Economy 99 (4), 689{721. Fama, E. (1985). What's dierent about banks? Journal of Monetary Economics 15, 29{39. Gande, A., M. Puri, and A. Saunders (1999). Bank entry, competition and the market for corporate securities underwriting. Journal of Financial Economics 54 (2). 22 See pp in Judd (1996). 31

32 Gande, A., M. Puri, A. Saunders, and I. Walter (1997). Bank underwriting of debt securities: Modern evidence. Review of Financial Studies 10 (4). Hansen, R. S. (September 1999). Do investment banks compete in IPOs?: the advent of the `7% plus contract'. mimeo. Helwege, J. and N. Liang (1996). Is there a pecking order? Evidence from a panel of IPO rms. Journal of Financial Economics 25, 429{458. Houston, J. and C. James (1996). Bank information monopolies and the mix of private and public debt claims. Journal of Finance 51, 1863{1889. James, C. (1987). Some evidence on the uniqueness of bank loans. Journal of Financial Economics 19, 217{235. Judd, K. L. (1996). Approximation, perturbation, and projection methods in economic analysis. In H. M. Amman, D. A. Kendrick, and J. Rust (Eds.), Handbook of Computational Economics, Volume I (Handbooks in Economics: 13). Amsterdam: Elsevier. Kanatas, G. and J. Qi (1998). Underwriting by commercial banks: Incentive conicts, scope economies, and project quality. Journal of Money, Credit, and Banking 30 (1), 119{133. Kroszner, R. and R. Rajan (1994). Is the Glass-Steagall Act justied? A study of the U.S. experience with universal banking before American Economic Review 84 (4), 810{832. Leland, H. and D. Pyle (1977). Information asymmetries, nancial structure and nancial intermediaries. Journal of Finance 32, 371{387. MacKie-Mason, J. K. et al. (1990). Do rms care who provides their nancing? In R. G. Hubbard (Ed.), Asymmetric Information, Corporate Finance, and Investment. Chicago and London: University of Chicago Press. A National Bureau of Economic Research Project Report. McFadden, D. (1981). Econometric models of probabilistic choice. In C. Manski and D. McFadden (Eds.), Structural Analysis of Discrete Data with Econometric Applications. MIT Press. McLachlan, G. J. and T. Krishnan (1997). The EM Algorithm and Extensions. New York: John Wiley and Sons, Inc. Petersen, M. A. and R. G. Rajan (1994, March). The benets of lending relationships: Evidence from small business data. Journal of Finance 49 (1), 3{37. 32

33 Petersen, M. A. and R. G. Rajan (1995, May). The eect of credit market competition on lending relationships. Quarterly Journal of Economics 110 (2), 407{43. Puri, M. (1994). The long-term default performance of bank underwritten security issues. Journal of Banking and Finance 18 (2), 397{418. Puri, M. (1996). Commercial banks in investment banking: Conict of interest or certication role? Journal of Financial Economics 40, 373{401. Puri, M. (1999). Commercial banks as underwriters: Implications for the going public process. Journal of Financial Economics 54 (2). Rajan, R. (1992a). A theory of the costs and benets of universal banking. mimeo., University of Chicago. Rajan, R. G. (1992b). Insiders and outsiders: The choice between informed and arm's-length debt. Journal of Finance 47 (4), 1367{400. Sudip Datta, M. I.-D. and A. Patel (1999). Bank monitoring and the pricing of corporate public debt. Journal of Financial Economics 51, 435{

34 Figure Price (%) Price Issue Amount ($ million)

35 Figure 2 Firm s Choice Set investment banks commercial banks level m upper nest underwriting bank underwriting bank level j lower nest

36 Underwriting Mechanism Syndicate Agreement 1. names of participants 2. Allotment to each member 3. responsibility of unsold securities Firm book runner (lead manager) Underwriting Agreement 1. price guaranteed to issuer 2. initial price offered to public 3. responsibility of unsold securities Underwriting Syndicate Marketing and distribution Investor Investor Investor Investor Investor Figure 3

37 table1.as Table 1 Sample Summary Statistics Bank issues Investment bank issues All issues Segment no. no.(%) volume ($m) no. no.(%) volume ($m) no. no.(%) volume ($m) Issue size($m) <= % $1, % $5, % $7, < <= % $13, % $68, % $82, < % $14, % $175, % $189, % $29, , % $250, , % $279, Maturity(yr.) <= % $3, % $29, % $32, < <= % $22, % $155, % $177, < 25 11% $3, % $65, % $69, % $29, , % $250, , % $279, Credit Rating Caa-Ba % $11, % $75, % $87, Baa1-Aaa % $17, % $174, % $192, % $29, , % $250, , % $279,921.10

38 table2-3.as Table 2 Sample Tabulation by previous issue experience no. of Bank issues Investment bank issues All issues Issuer issues(%) no. no.(%) volume($m) volume(%) no. no.(%) volume($m) volume(%) no. no.(%) volume($m) volume(%) First-time 44% % $16,447 13% % $107,909 87% % $124, % Seasoned 56% % $12,745 8% % $142,820 92% % $155, % Total 100% % $29,192 10% % $250,729 90% % $279, % Table 3 Sample Tabulation by SIC codes Issuer s no. of Bank issues Investment bank issues All issues SIC issues(%) no. no.(%) volume($m) volume(%) no. no.(%) volume($m) volume(%) no. no.(%) volume($m) volume(%) 000 s 1% 1 11% $300 18% 8 89% $1,335 82% 9 100% $1, % 1000 s 13% 30 15% $4,015 12% % $28,356 88% % $32, % 2000 s 30% 61 13% $8,270 9% % $80,577 91% % $88, % 3000 s 22% 59 18% $7,006 10% % $62,465 90% % $69, % 5000 s 15% 34 15% $5,375 14% % $34,173 86% % $39, % 7000 s 15% 31 13% $3,421 10% % $32,429 90% % $35, % 8000 s 4% 7 11% $805 7% 58 89% $11,395 93% % $12, % Total 100% % $29,192 10% % $250,729 90% % $279, %

39 Table 4: Estimation Results of Basic Model Demand estimates Reputation category (none) Variable names PRICE LOAN Variables a b Estimates Standard errors (0.0454) (0.1789) Inclusive Variable names Value FIRST-TIME AMOUNT Variables λ δ1 δ2 Estimates Standard errors (0.0924) (0.1402) (0.0802) Price estimates function of function of Past MTN (shelf LOAN Variable names MATURITY JUNK BOND Issue Experience registration) (junk only) commercial banks Bank Bank Bank Bank Bank investment banks Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank

40 Table 5: Estimation Results Demand estimates Reputation category Non-junk issuers Junk bond issuers Variable names PRICE LOAN PRICE LOAN Variables a1 b1 a2 b2 Estimates Standard errors (0.0969) (0.0424) (0.2277) (0.3158) Inclusive Variable names Value FIRST-TIME AMOUNT Variables λ δ1 δ2 Estimates Standard errors (0.1375) (0.1418) (0.0755) Price estimates function of function of Past MTN (shelf LOAN Variable names MATURITY JUNK BOND Issue Experience registration) (junk only) commercial banks Bank Bank Bank Bank Bank investment banks Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank

41 Table 6: Estimation Results Demand estimates Reputation category Non-first-time issuers First-time issuers Variable names PRICE LOAN PRICE LOAN Variables a1 b1 a2 b2 Estimates Standard errors (0.0745) (0.2412) (0.0514) (0.2754) Inclusive Variable names Value FIRST-TIME AMOUNT Variables λ δ1 δ2 Estimates Standard errors (0.1411) (0.1446) (0.0770) Price estimates function of function of Past MTN (shelf LOAN Variable names MATURITY JUNK BOND Issue Experience registration) (junk only) commercial banks Bank Bank Bank Bank Bank investment banks Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank

42 Table 7: Effect of Increase in Price (mean own-price elasticities) Reputation category All issuers Non-junk issuers Junk bond issuers commercial banks Bank Bank Bank Bank Bank Mean Std. Dev investment banks Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Mean Std. Dev

43 Table 8: Effect of Increase in Price (% change in Own Total Demand from 1% increase in price across sample) Reputation category All issuers Non-junk issuers Junk bond issuers commercial banks Bank Bank Bank Bank Bank Mean Std. Dev investment banks Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Mean Std. Dev