Relationship Banking and the Pricing of Financial Services

Appendices To Accompany

“Relationship Banking and the Pricing of Financial Services,”

Journal of Financial Services Research, June 2009

By Charles W. Calomiris and Thananvut Pornrojnangkool

This document includes three appendices: Appendix A2: Comparison Appendix, which explains the factors that account for differences between our findings and those of Drucker and Puri (2005), Appendix A3: Two-stage Least Squares Regressions, which present results using two-stage least squares rather than GMM estimation, and Appendix A4: Loan Regressions Robustness, which performs several tests for the robustness of our loan regression results.

Appendix A2

Comparison Appendix

Our paper differs from Drucker and Puri (2005) (hereinafter DP) in several respects. The purpose of this Comparison Appendix is to describe differences in data, assumptions, and methodologies, and review which differences are most important in explaining the differences in our findings with regard to loan pricing. The differences discussed here include: (1) econometric approach, (2) our approach to controlling for differences among firms, (3) definitions of what constitutes a relationship banker, (4) our treatment of differences in relationship pricing related to the sequencing of transactions, (5) differences in sample periods, (6) differences in types of loans included, (7) the functional form chosen for loan supply (that is, log of credit spread vs. credit spread as dependent variable), and (8) structural or non-structural approach to modeling the lending market. To summarize our discussion below, the first four of these eight differences seem to be the most important for explaining how our loan pricing results differ from DP.Controlling for risk differences and taking sequencing into accountare particularly important differences. With respect to underwriting costs, differences in the definition of what constitutes a relationship banker seem to be more important in explaining differences between their results and ours.

Loan Pricing: Econometric Approach

In our analysis of loan pricing, we differ from Drucker and Puri (2005) in econometric approach. They forecast matching (relationship bundling of loans and equity underwriting) using a probit model, which includes a numerical credit ratings scale, loan size, loan type, loan maturity, and year and industry dummies. They then group each matched observation with a set of non-matched neighbors that have the closest forecasted matching score from the probit, and then calculate the average spread differences between matched loans and their non-matched neighbors. This approach ignores the effects of risk and other firm factors on loan pricing unrelated to matching. Indeed, we believe that the forecasters they employ are not proper instruments for forecasting matching, since they are variables likely to be related to loan pricing irrespective of matching. We are unable to identify valid instruments, and therefore, do not attempt to forecast matching in our analysis.

In Table A2-1 we approximately replicate the DP approach, under our definitions of what constitute banking relationships in lending and underwriting, which differ somewhat from those of DP (as discussed further below). Despite the differences in our definitions of relationships, using their sample period and using their specification of temporal proximity (six months or less between lending and underwriting events) we are able to reproduce their central finding for loan pricing: matched loans provided by commercial banks are significantly cheaper than other loans. This finding, however, is not very robust. When one expands the sample to our larger sample period, junk-rated deals (by both commercial banks and investment banks) have statistically significantly lower loan pricing, but underpricing by commercial banks, per se, declines in magnitude and statistical significance. When one uses a 12-month period in the definition of temporal proximity, results are further weakened, and the only statistically significant results are for our sample period, for the sub-sample of junk-rated deals.

In Table A2-2 we explore whether, within the DP framework, the ordering of loans and underwriting (i.e., whether underwriting precedes or follows lending) matters for pricing. We find no robust difference between MSE or MPE transactions in loan spread differences between relationship bundlers and their near neighbors, as defined in their approach. Differences across related cells between MSE and MPE sub-samples in Table A2-2 are not robust to differences in sample periods or the definitions of temporal proximity of equity underwriting and loan transactions (six months vs. twelve months).

In Table A2-3 we examine the effect on the DP findings of controlling for the differences in risk and other loan attributes when comparing matched loans with their “neighbors.” We find that the negative loan pricing effects identified by DP are not present once one controls for risk using our vector of control variables (constructing residual spreads using a simple OLS model predicts spreads using controls for firm characteristics, transaction characteristics, sequencing controls like PE, PD, SD and SE, and loan amount). Furthermore, as shown in Table A2-4, once one controls for risk, the sequencing of underwriting and lending matters for loan pricing, and the results are broadly consistent with our findings. That is, MSE matching tends to be associated with positive spreads, and MPE matching tends to be associated with negative spreads. We conclude that DP’s findings differ from ours primarily because they do not control for differences in loan attributes unrelated to matching.

What Is a Relationship Banker?

In our study, we deem a bank to have a lending relationship with a client if the bank is listed at the agent level as a member of the loan syndicate; DP, in contrast, define all participants as having a relationship with the borrower. We deem a bank to have an underwriting relationship with a client if the bank is the book runner or joint book runner for the underwriting (of which there are usually only one or two); DP define only the first name in the underwriting list as having the underwriting relationship. In our view, DP’s inclusion of all loan participants creates too broad a definition of lending relationships, and its exclusion of joint book runners creates too exclusive a definition of underwriting relationships. We note that most studies of relationship banking with which we are familiar follow the convention of assuming that relationships adhere to loan originators and agents, but not to participants. It would be difficult to replicate the DP definitions of relationships for our sample, so we did not do so. In any case, differences in the definition of relationships does not explain the differences in our key findings, since we were able in Table A2-1 to replicate DP’s main results using our definition of relationships and their methodology, and to show that those results are not robust, within their methodological framework, to the inclusion of our controls.

In the remainder of this Comparison Appendix, we review how other differences in specification between DP and our study affect our results.In Tables A2-5 to A2-7 we report various models that show the effects of varying particular assumptions. Table A2-5 to A2-7 contain simple, non-structural models of loan spreads with some basic controls, but varying specifications using different matching periods, sample periods, and log vs. non-log definitions of the dependent variable. The last four sets of columns in Table A2-7 are true reduced-form regressions, which measure the effects of all exogenous variables in our Loan Supply and Loan Demand equations on loan spreads. The key findings, that the average effect of matching is positive (CONCURE), that MPE is negative, and that MSE is positive, is robust across all the specifications.

Thus, within our methodological approach, our key results for loan pricing are robust to a variety of specification and sample variations. Neither is the structural modeling of the loan market (identifying supply and demand) crucial for those conclusions. Inclusion of controls in the regression, however, are important; the robust statistical significance of the MSE and MPE effects shown in Tables A2-5 to A2-7would not hold in the absence of controls.

Equity Underwriting Costs

Our results for equity underwriting costs, reported above, also differ from those of DP. We report additional regression results for comparison purposes in Tables A2-8 through A2-10. DP use a different definition of a lending relationship and also include variables that capture distant prior lending actions in their analysis of equity underwriting costs. They also use a different functional specification of the relationship between equity underwriting cost and the size of the offering. Table A2-8 comes close to replicating their approach, although it differs from DP because it uses our definitions of relationships and it does not capture distant prior lending activities. We are unable to replicate their central finding that stand-alone investment banks charge less on underwriting equity when equity underwriting is part of a relationship in which underwriting and lending both occur. In fact, we tend to find a positive effect from the existence of a lending relationship, similarly to what we found in our equity underwriting regressions in Table VII. Table A2-9 breaks down the concurrent loan relationship dummy into MPL and MSL according to our definitions of these variables and we find similar results to those in Table A2-8.

These tables also divide the sample into investment grade or non-investment grade subsamples, which produces a further difference from DP; we find a (non-robustly) negative relationship effect of concurrent loans for equity underwriting costs of investment grade issuers, while DP’s result of relationship discounts on equity underwriting costs was driven by non-investment grade firms.

In Table A2-10, we replace the DP control variables with our control variables and are able to confirm our findings in Table VII of the main text. We conclude that fundamental differences between our definition and their definition of a banking relationship (their assumption that a bank’s involvement in a loan participation is a sufficient criterion for establishing a relationship relevant for the pricing of underwriting services, and the tracking of loan transactions from the distant past) seem to be the source of differences between our findings and theirs. As we have discussed above and in the body of the paper, we believe that our definition of a relationship is more appropriate on a priori grounds.

1

1

Appendix A3

Two-stage Least Squares Regressions

The results from two-stage least square estimates of Loan Supply and Loan Demand equations are similar to those of GMM. At the bottom of the table A3-1, we display results for tests of the significance of these instruments in the first-stage regression of LNAMT on all exogenous variables. Individually and jointly, these instruments are correlated with LNAMT. We also implement a regression-based test of the Hausman (1978) procedure to test the null hypothesis that instruments are exogenous. As shown in the overidentification tests in the table, the value of the test statistic is 0.37 for model A and 0.26 for model B, indicating that one cannot reject the null of instruments exogeneity.

In addition, we utilize the instruments to test for the endogeneity of LNAMT in the spread regression. If LNAMT is exogenous in the spread equation, then ordinary least squares and two-stage least squares estimates of all coefficients should differ only by sampling error. The test is implemented by first regressing LNAMT on all exogenous variables to obtain its residual. Then, we add the residuals from this regression to the spread regression (1) to obtain the OLS estimate. The t-statistic of the residual term in this augmented spread regression can be used as a test statistic for the null hypothesis that LNAMT is exogenous.[1] Our t statistic has a value of -12.77 for model A, and -11.51 for model B. Thus, we clearly reject the null hypothesis that LNAMT is exogenous.

Two-stage least squares estimates of the Loan Demand equation are presented in models C to E in Table A3-2. The sign of LNSPREAD is negative and significant, confirming the demand interpretation of the equation. Our instruments for LNSPREAD also work well in all specification tests reported. In the first-stage regression, PRIME and SIC2LESH are very significant instruments in explaining LNSPREAD, as shown in the test statistics at the bottom of the table. In addition, we can not reject the null hypothesis that our instruments are exogenous in the overidentification test, which validates our choices of instruments.

1

1

Appendix A4

Loan Regressions Robustness

Robustness Tests

We perform several tests for the robustness of our loan regressions. First, we exclude commercial paper backups and term loans from our loan regressions (i.e., we include only revolvers). The results are similar to the previous results and are shown in Table A4-1. Second, we exclude commercial paper backup loans altogether from the original DealScan dataset when we define financing windows and matched loans. The results for this restricted sample are qualitatively similar to the full sample and revolver loans results, and thus we do not show them here. Additionally, we estimate loan demand and supply equations allowing IBs and universal banks to have different demand and supply schedules (i.e., allowing endogenous variables to interact with the IB dummy). The interaction terms are not significant, thus we do not explore this specification further.

Sub-Samples Analysis

Table A4-2 presents the GMM estimates of spread regressions broken down by time and borrower sales.[2] The evidence of a loan pricing premium on loans that precede matched equity transactions holds when we split our sample pre- and post-1998, but results are more significant post-1998, which corresponds to the period when the Glass-Steagall Act was no longer in effect. This pricing premium applies across the borrower size spectrum, although the coefficients for MSE are not significant for the smallest size borrowers (with less than $250 million in annual sales) and the largest size borrowers (with more than $10 billion in annual sales). This finding is consistent with our previous interpretation that the MSE premium reflects bank quasi rent extraction by virtue of their relationships. In our sample, there are few MSE transactions for the smallest size category of borrowers, which can explain the larger standard errors for that coefficient. For the largest borrowers, we hypothesize that the underwriting market is highly competitive (i.e., lenders lack significant private information about these borrowers) so that banks seeking to exploit their relationships to extract quasi rents (as in Rajan 1992) would fail because they have no market power in the lending market.

The sign for the coefficient of IB*MPE is consistently negative pre- and post-1998 and across borrower sizes. However, the coefficients are significant only for the post-1998 period and for large borrowers (those with annual sales more than $1 billion). This finding is also consistent with the hypothesis that investment banks suffer cost disadvantages relative to universal banks in providing loans, and therefore, are forced to compete in the loan market by providing “rebates” of their underwriting fees in the form of pricing discounts for loans that follow equity offerings. Since this “rebate” is costly, it is logical for them to offer it only when they have to do so (i.e., on deals where revenue is large, and for which the competition from universal banks is strongest – namely loans to large borrowers).

1

1

[1] See for example p.118 of Wooldridge (2002) for explanation of regression-based approach to endogeneity test.

[2] The results are similar for the two-stage least squares estimators.