Bank Capital and Lending: Evidence from Syndicated Loans

Transcription

1 Bank Capital and Lending: Evidence from Syndicated Loans Yongqiang Chu, Donghang Zhang, and Yijia Zhao This Version: June, 2014 Abstract Using a large sample of bank-loan-borrower matched dataset of individual loans, we find that, conditional on loan demand, the lead bank of a syndicated loan contributes more to the loan if it has a higher capital ratio. The result holds if within-loan estimations are employed to further remove the impact of demand side factors. Additionally, using the Troubled Asset Relief Program (TARP) as a quasi-natural experiment, we find that TARP recipients increase their contributions to syndicated loans after receiving TARP capital injections. Taken together, we provide new evidence on the causal effect bank capital on lending. Keywords: Bank Capital, Syndicated Loans, Lending, TARP All authors are from Moore School of Business at the University of South Carolina, 1705 College Street, Columbia, SC Chu can be reached at yongqiang.chu@moore.sc.edu, Zhang can be reached at zhang@moore.sc.edu, and Zhao can be reached at zhao45@ .sc.edu. The authors thank Allen Berger for his comments and Allen Berger and Raluca Roman for sharing their data on the TARP program.

2 Both regulators and academics are interested in how bank capital affects bank lending. This issue is particularly acute in light of debates over the potential economic consequences of the strengthened bank capital requirements proposed in the aftermath of the recent financial crisis. Both Basel III and the Dodd-Frank Act propose to substantially increase bank capital requirements on the premise that more capital can provide the safety net and thus increase financial stability and efficiency. However, bankers have been arguing that raising capital requirements could impede lending. Although the academic literature has devoted much attention to understanding this important issue, a clear identification of a causal effect from bank capital to bank lending has proven to be difficult, especially with only aggregate or bank balance sheet data. The main obstacle is to separate the effect of bank capital on lending from (often unobservable) demand side factors because demand side factors can be correlated with bank capital. In this paper, we make new attempts to uncover how capital affects lending using a dataset that matches individual loan-level data (from LPC DealScan) and borrower financial statement data (from Compustat) with bank s balance sheet data (from the Consolidated Report of Condition and Income or the Call Report ). Specifically, we study how a bank s capital level affects its lending measured by its allocation share in syndicated loans. We examine all banks that participate in a syndicated loan, including both the lead bank that arranges the loan and the participant banks that often just provide funding for the loan. A bank s allocation share for a syndicated loan, which is also referred to as a bank s allocation or a bank s share, is the ratio of the funds contributed by the bank over the total loan amount. Our empirical approach enables us to better separate the effect of bank capital on lending from the impact of demand side factors, therefore providing a clean identification of the causal effect of bank capital on credit supply. We rely on detailed loan level data rather than aggregate or bank balance sheet data to identify bank lending behavior. This renders us two advantages compared with previous studies. First, we are able to observe a bank s allocation, or share, within a loan, which has been conditioned on the total amount of credit demanded by the borrower (Duchin and Sosyura (2014)). So any effect of capital 1

3 on the bank s allocation decision within a loan is more likely to be driven by lender (bank) side factors. Second, because we link each loan to its borrower, we are also able to control for a broad set of borrower-specific variables. This can further mitigate the potential bias caused by the correlation between omitted demand side factors with bank capital. Using a large sample of 4,356 (5,634) syndicated loan packages (facilities) made to U.S. firms between 1996 and 2012, we find that lead banks allocation shares within loans are positively related to the lead banks capital ratios. Specifically, in our baseline OLS regressions that control for a variety of loan, borrower and lead bank characteristics, we show that one standard deviation increase of the lead bank s total capital ratio causes an increase of the lead bank share by 3.2 to 3.5 percent. This result persists at both the package and the facility levels. The positive effect of lead bank capital on lead bank share remains intact if we include bank fixed effects to control for unobserved bank characteristics. It is also robust to the inclusion of state-year fixed effects that account for the possible impact of borrower local economic conditions. We then employ two additional identification strategies to further establish the causal effect of bank capital on lending. First, we examine how the differences of capital levels of banks participated in the same loan affect their contributions to the loan, which we call the within-loan estimation. Empirically, we regress a bank s share in a package/facility on the bank s capital ratio and other characteristics with a dummy variable for the package/facility. Note that the estimation is executed by first washing out the package/facility fixed effects, which also differences out any possible confounding firm and loan-level factors that are otherwise unobservable. That is, by estimating the effect of bank capital on the within-loan difference in lender shares, we further remove the impacts of demand side factors. The point estimates of the impact of a bank s capital ratio on the bank s share at the package level range from 0.51 to 0.63, and at the facility level, such estimates range from 0.58 to These point estimates are about half of those for across different loans and are generally statistically significant. These results provide further support that a higher capital ratio of a bank leads to more lending. 2

4 In the second identification strategy, we exploit plausibly exogenous variations in bank capital levels generated by the Troubled Asset Relief Program (TARP). Using the withinloan estimation, we estimate the effect of TARP capital injection in a difference-in-difference framework and find that TARP recipient banks contribute more to syndicated loans than non-tarp banks after they receive TARP funding. More specifically, one percentage increase in a bank s capital due to TARP capital injection would result into 2.29/2.17 percent (at the facility level) or 3.69/3.60 percent (at the package level) higher share of lending from the recipient bank. This finding indicates that, conditioning on the overall loan amount, an exogenous increase in capital is associated with an increase in a bank s contribution to a loan. This result again confirms a casual effect of bank capital on lending. Note that the exercise using TARP helps us deal with the more general endogeneity problem, which can be caused not only by demand side factors but also by any other omitted variables, or other problems. With our main findings described above and others that will be discussed later, our paper contributes to the large literature on how bank capital affects bank lending. Many theories suggest that bank capital is positively related to bank lending. First, low bank capital increases the risk premium demanded by depositors and therefore increases the cost of bank loans (Jayaratne and Morgan(2000), Kishan and Opiela (2000), and Van den Heuvel (2002)). Second, banks with lower capital may cut back lending to shrink their balance sheets. Third, under risk-based capital regulations, banks may simply substitute loans, which usually have higher risk weights, for safer assets, especially when bank capitals are low. However, Diamond and Rajan (2000) make an opposite argument that low levels of bank capital create the incentive for banks to monitor and therefore should increase lending. Early empirical literature on the relationship between bank capital and lending has mostly focused on the causes and consequences of the credit crunch in the late 1980 s and early 1990 s and the U.S. adoption of the Basel Accord in the early 1990 s. Most early literature finds a significant impact of bank capital on lending. For example, Bernanke et al. (1991) use state-level data and find that one percentage point increase of the capital ratio leads to an 3

5 increase of the loan growth rate by 2.6 percentage points. Hancock and Wilcox (1993) find that one dollar bank capital shortage (relative to regulatory capital) leads to a reduction of bank credit by three dollars. Brinkmann and Horvitz (1995) find that banks with a higher risk-based capital ratio have substantially higher loan growth rates during the early 1990 s credit crunch. One exception is Berger and Udell (1994), who find very limited effects of the risk-based capital ratio on loan growth rates using bank-level data from the Call Report. While the early empirical literature provides important insights into the effects of bank capital on bank lending, most papers have difficulties in separating the effect of bank capital from that of demand side factors. More recent literature has devoted much attention to separating the effect of bank capital on bank lending from that of the demand side factors. For example, Peek and Rosengren (1997)usethedramaticdeclineoftheJapanesestockmarketinthelate1980sandearly1990s as an exogenous shock to capitals of Japanese banks and study how decreased capital levels of U.S. branches of Japanese banks affect their lending. They find that these subsidiaries substantially reduced their commercial lending in the U.S. Peek and Rosengren(2000) use the same exogenous shock and further demonstrate that loan supply shortage due to constrained bank capital has real effects on economic activities. Puri et al. (2011) use German banks exposure to the U.S. subprime market as an exogenous shock to capitals of German banks and find that German banks exposed to the U.S. financial crisis substantially reduce lending. Most recently, Rice and Rose(2012) use the bailout of the Government Sponsored Enterprises (GSEs, or specifically Fannie Mae and Freddie Mac) as a natural experiment, because many U.S. banks held a substantial amount of preferred equity of the GSEs, which was wiped out during the bailout. They find that banks more exposed to the GSEs experienced substantial decreases of capital and reduced lending after the bailout of Fannie Mae and Freddie Mac. With the exception of Puri et al. (2011), all other aforementioned papers still rely only on aggregate data or bank balance sheet data, and therefore are still unable to fully control individual borrower characteristics that may affect loan demand. In this paper, by focusing on the effect of bank capital on banks allocation shares in 4

6 syndicated loans, we also attempt to separate the effect of bank capital from demand side factors. Our tests provide new insights because we study the relative movements conditional on the total amount of credit. This is particular true for the within-loan estimations with only the participant (non-lead) banks since these participant banks are unlikely to have much influence on the determinations of the total loan size. So the fact that a participant bank with more capital contributes more to a given loan provides a clean identification of the impact of bank capital on lending. 1 The positive and causal effect of bank capital on syndicated loan shares found in our paper thus suggests that higher levels of bank capital may contribute to more overall lending activities. Our paper also contributes to the literature on loan syndicate structure. The existing literature has focused on how asymmetric information between the lender and the borrower or asymmetric information between lead banks and participant banks affects syndicate structure. Sufi (2007) finds that lead banks retain larger shares of syndicated loans and form more concentrated syndicates when borrowers require more intense monitoring and due diligence. Ivashina (2009) finds that asymmetric information between lead banks and participating banks also leads to higher shares retained by lead banks. Gatev and Strahan (2009) examine how bank liquidity affects commercial banks shares in syndicated loans. Ivashina and Scharfstein (2010) study how syndicate structure varies over the business cycle. To the best of our knowledge, this paper is the first to study how supply side factors affect syndicate structure. More importantly, studying the effect of bank capital on syndicate structure enables us to better understand how bank capital levels affect real economic activities. As capital affects the lead bank s allocation share in syndicated lending, the lending share can in turn affect the loan spreads that other participating banks may require (Ivashina (2009)). Therefore, the effect of bank capital on syndicate structure establishes a link between bank capital and the cost of external financing for the borrowers, which can then affect the borrowers real activities. 1 In this regard, our approach is similar to Kashyap et al. (1993), who use the relative movements between bank loans and commercial papers to identify the transmission of monetary policies through the lending channel. 5

7 The rest of the paper is organized as follows. Section I describes the data. Section II presents the baseline results on bank capital and lending. Section III presents results from alternative identification strategies. Section IV presents the results on bank capital and syndicate size. Section V concludes. I. Data and Descriptive Statistics A. Sample Construction and Key Dependent Variables Our sample construction begins with a sample of 222,991 distinct loan facilities between January 1996 and December 2012 from LPC Dealscan. We begin our sample in 1996 because only since then do banks report their risk-based capital ratios in the Call Report. These 222,991 facilities belong to 155,345 distinct deal packages. 2 We then use the DealScan- Compustat link file provided by Chava and Roberts (2008, updated in August 2012) to match the loan sample with borrowers financial statement information from Compustat. The match produces a sample of 97,924 facilities (72,258 deal packages) for which the borrowers financial information can be found in Compustat. The key dependent variable of our empirical analysis is a lender s allocation of a loan, which is referred to as Bank Share. 3 The DealScan database reports lender identities and their loan allocations at the facility level. Our unit of analysis is at both the facility and the deal package levels. To calculate a bank s allocation in a package, we first obtain the bank s allocations in all facilities within the package and then aggregate these allocations at the package level using individual facility amounts and the total package amount. For example, if a bank participates in both two facilities in a deal package and the two facility amounts are 60% and 40% of the total package amount, and if the bank contributes 30 percent in the first facility and 50 percent in the other facility, we calculate the bank s share in the entire deal package as: 60% 30%+40% 50% = 38%. 2 A deal package can contain multiple facilities and each facility within a package can be of the same or different types of credit. 3 or Lead Bank Share for lead banks 6

8 We calculate bank shares for both lead banks and participant banks. The DealScan database reports the roles of lenders in each facility. We follow Ivashina (2009) to identify the lead bank(s) of a facility. If a lender is reported as the administrative agent, it will be defined as the lead bank. If no lender is reported as the administrative agent, we define lender(s) who act as the agent, arranger, book-runner, lead arranger, lead bank, or lead manager as the lead bank(s). A lead bank of any facility in the package will be regarded as the lead bank of the package. Among the 72,258 deal packages (97,924 facilities) for which the borrowers financial information can be found on Compustat, we are able to calculate bank shares at the package level for 31,113 deal packages (38,968 facilities) between 1996 and We link loans in DealScan to bank information in the Call Report. Because there is no common identifier across the two databases, we use a text matching program to match bank names reported in DealScan with bank legal names in the Call Report. Wherever possible, we also use a bank s geographical information (city and/or state) reported in both databases to facilitate the matching. We then manually check all the automated matching results to ensure matching accuracy. We also rely on information provided by the FDIC institution search ( and/or by the National Information Center (NIC) to identify DealScan lenders that are not matched by computer programs. For all lenders in the DealScan universe, we are able to identify 1,269 unique U.S. financial institutions that have Call Report information. Note that the 1,269 financial institutions include lead and participating banks in the entire DealScan universe, not just our sample. For the sample of 31,113 packages with valid lead bank share information, we then match 4 The reduction of the sample size is mainly due to the fact that the lender allocation information is entirely missing for about 72% of all facilities in the original DealScan database (also see Ivashina (2009)). There are also cases in which shares of only some (but not all) lenders of a facility are reported. To ensure accuracy, we exclude packages with facilities that have missing information on any lender shares. The only exception is when a package has only one lender and its allocation information is missing. In this case, we set the lender share to be 100%. We also exclude packages with incorrect lender share information (e.g., those packages in which the sum of all lender shares are more than 101% based on DealScan s original information - we choose 101% because some small rounding errors could lead to the summation of all lender shares to slightly exceed 100%). We are indeed able to calculate shares of all lenders for 34,231 packages between 1996 and However, we cannot reliably identify any lead bank among all lenders based on Ivashina (2009) s lead bank identifying method for the other 3,118 (34,231-31,113) packages. 7

9 them with borrower characteristics from Compustat and bank characteristics from the Call Report. In particular, borrower characteristics are obtained from the Compustat Annual database as of the fiscal year ending immediately prior to the deal activation date. Bank characteristics are obtained from the Call Report as of the filing in the calendar quarter immediately prior to the deal activation date. We only keep loan packages made by U.S. commercial banks. We also exclude loan packages made to regulated and financial borrowing firms (two digits SIC code equals to 49 or between 60 and 69) and those to non-u.s. firms. These requirements reduce our sample to 13,183 packages. 5 The requirement of simultaneously having key firm characteristics from the Compustat and key bank characteristics variables from the Call Report further reduces the sample to 4,772 unique loan packages. Finally, we only keep packages that contain credit lines, or term loans, or both for subsequent analysis. This leads to a final sample of 4,356 unique deal packages, our main sample at the package level. We focus on credit lines and terms loans because we are interested in how bank capital affects credit supply to non-financial corporate borrowers and credit lines and terms loans are the most popular types of bank financing obtained by non-financial firms (see, e.g., Rauh and Sufi (2012) and Colla et al. (2013)). Our main results are robust if we include other infrequent packages such as bridge loans, standby letters of credit, and leases. Our final sample involves 2,435 unique borrowing firms and 235 unique lead banks. Among the 4,356 packages, 4,297 packages have only one lead bank, 56 packages have two lead banks, one package has three lead banks, and two packages have four lead banks. Our baseline analysis focuses on lead banks since they originate and are in a management position of a deal package. In this case, the unit of observation is a loan-lead bank pair. Therefore our main sample at the package level has 4,420 individual observations (package-lead bank pairs). The 4,356 packages correspond to 5,634 facilities which constitute a facility level sample of 5,698 individual observations (facility-lead bank pairs). We report the detailed decomposition of the 4,356 loan packages according to loan types in Panel A of Table I. Packages that contain only credit lines and both credit lines and terms 5 The large reduction in sample size in this step results primarily from the exclusion of loans to non-u.s. firms. 8

10 loans take a significant portion (93.89%) of our whole sample. This observation is consistent with existing literature suggesting that credit line is an instrumental component of corporate external finance (e.g., Sufi (2009), Campello et al. (2011)). Panel B of Table I shows the top ten lead banks with the largest number of deals led by them. Not surprisingly, giant banks such as JP Morgran Chase, Bank of America, and Wells Fargo appear on this list. The top ten most active banks together lead about 67.97% of all packages in our sample. Because of multiple deals led by these banks, we are able to estimate a lead bank fixed effects model in addition to the simple OLS model to account for any unobservable time invariant bank characteristics. B. Independent Variables In the following subsections, we discuss explanatory variables used in our regressions. The detailed variable definitions can be found in Appendix A1. B.1. Bank Level Explanatory Variables The key independent variable of interest is a bank s capital ratio. We use three different measures of bank capital. Our first and main measure, Total Capital Ratio, is defined as total capital divided by bank total risk-weighted assets. Our second capital measure is Tier 1 Capital Ratio, which is defined as tier 1 capital divided by bank total risk weighted assets. Compared with tier 2 capital, tier 1 capital measures the core capital that is not subject to withdrawal by equity holders. Our third measure is Leverage Ratio, which is defined as tier 1 capital divided by bank total (un-weighted) assets. The Leverage Ratio is a non-risk-based capital measure. We use Leverage Ratio to make sure that our results are not driven by a bank s incentive to strategically manage its risk-based assets. We include a broad set of bank level control variables in our regressions. These control variables are defined as follows. Log (Bank Total Assets): defined as the natural logarithm of bank total assets ($thousand). This is a measure of bank size. Log (Bank Total Assets) can have a mechanical effect 9

11 on Lead Bank Share, as larger banks are able to lend more. Meanwhile, larger banks have greater capacities to absorb capital shocks because they tend to be better diversified and/or are more likely to have access to cheaper funds. We therefore expect Log (Bank Total Assets) to have a positive effect on lender shares. Bank Liquidity: defined as the sum of cash and available-for-sale securities divided by bank total assets. Banks with more liquid assets are more likely to be able to fund loans on the margin. As documented in Khwaja and Mian (2008), bank liquidity shortage can translate into loan declines of the borrowing firm. We therefore expect Bank Liquidity to have a positive effect on lender shares. Bank ROA: defined as bank operating income divided by total assets. More profitable banks are likely to have access to lower cost funds thus are at better positions to absorb negative capital shocks. We therefore expect Bank ROA to have a positive effect on lender shares. Loan Charge-Offs: defined as the total charge-offs on loans and leases divided by bank total assets. It measures a bank s loss incurred on previous loans it made (Santos (2011)). Murfin (2012) shows that banks write contracts with tighter covenants after suffering from payment defaults to their own loan portfolios. He argues that recent defaults update the lender s perception of its own screening ability. In this context, large losses could also change a bank s perception of its own screening ability, which could then affect the bank s decision to participate in subsequent loans. Overall, we expect Loan Charge-Offs to have a negative effect on lender shares. Loan Loss Allowance: defined as the total allowance for loan and lease losses divided by bank total assets. It reflects the lender s view on the future performance of its loan portfolio and its expectations of future market conditions. On one hand, Loan Loss Allowance reflects the quality of loans made in the past. On the other hand, Loan Loss Allowance is also correlated with future charge-offs. We thus expect Loan Loss Allowance to have a negative effect on lender shares. Risk-Weighted Assets: defined as total risk weighted assets divided by bank total assets. 10

12 It measures the overall riskiness of a bank s existing assets. In itself, higher risk should cause the bank to lend less. However, risk-weighted assets are also the denominator of riskbased capital ratios, and it remains to be tested whether risk-weighted assets should have additional effects other than its effect on risk-based capital ratios. In addition, we also include two measures of the bank s liability structure. Subordinated Debt: defined as total subordinated debt divided by bank total assets. On one hand, subordinated debt may act as a substitute for bank equity capital and its existence may indicate that the bank has access to public debt market (Santos (2011)). On the other hand, the literature argues that investors of subordinate debt can monitor and discipline bank to lower asset risk (Chen and Hasan (2011)). Deposits: defined as total deposits divided by bank total assets. Given that deposits are considered to be a stable and low-cost source of funding for banks, we expect it will have a positive effect on lender shares. Finally, we also include an indicator for bank holding companies, Bank Holding Company (BHC) Dummy, which equals oneif a bankis controlled by a bank holding company and zero otherwise. Banks controlled by bank holding companies may get support from the holding company or other institutions under the same holding company, and therefore may exhibit different lending behavior (Ashcraft (2008)). B.2. Borrower Characteristics We also include a number of borrower characteristics as regressors. We use these borrower characteristics to partially control for a borrower s demand for credit, which is impossible in most previous studies using only bank balance sheet information. These firm level controls can also capture the asymmetric information effect on lead banks share. Asymmetric information affects lead bank allocation share because higher lead bank allocation share can mitigate the potential moral hazard and adverse selection problems associated with syndicated lending (see e.g., Sufi (2007), Ivashina (2009)). Log (Firm Total Assets): defined as the natural logarithm of a firm s total assets ($mil- 11

13 lion). Mechanically, a lead bank may not be able to contribute a large portion to a large borrower s loan. A larger firm also implies less information asymmetry so that the lead bank may not have to hold a large fraction to keep skin in the game. So all else being equal, we expect Log (Firm Total Assets) to have a negative effect on lead bank allocation. Tobin s Q: defined as the market value of assets divided by book value of assets. High Q firms have more growth opportunities. Because growth options are difficult to value, it is often argued that high growth firms have more asymmetric information (e.g., McLaughlin et al. (1998) and Ongena and Smith (2001)). Tangibility: defined as total property, plant, and equipment divided by total assets. Tangible assets can serve as collateral, and therefore reduces the need of screening and monitoring (e.g., Barth et al. (2001) and Leary and Roberts (2010)). We therefore expect Tangibility to have a negative effect on Lead Bank Share. Asset tangibility can also be seen as a proxy for a borrower s debt capacity, and thus can be positively correlated with the borrower s demand for external credit. R&D: defined as R&D expenses divided by total assets. Firms with higher R&D expenses are more difficult to value (e.g., Aboody and Lev (2000) and Officer et al. (2009)). Therefore, high R&D firms are likely to have greater information asymmetry. Cash Flow Volatility: defined as the standard deviation of quarterly cash flows calculated over the last three years. High cash flow volatility implies higher degree of information asymmetry. Firms with more volatile cash flows are also considered to be riskier by creditors (Sufi (2009)). Leverage: defined as the total debt divided by book value of assets. Profitability: defined as the operating income before depreciation divided by total assets. Cash Holdings: defined as cash and marketable securities divided by total assets. Rated Dummy: an indicator variable, which equals one if the firm has an S&P long-term credit rating and zero otherwise. To sum up, including firm-level variables allows us to explicitly control for the effect of demand side factors, which is often not possible in studies relying only on bank balance sheet 12

14 data. C. Summary Statistics We present the summary statistics for the key variables on loan, bank, and borrower characteristics in Table II. Panel A of Table II reports the descriptive statistics of key loan characteristics at both the package and the facility levels. The sample is the 4,356 deal packages (5,634 facilities). Note that the number of observations is greater due to multiple lead or participant banks per package/facility. On average, a lead bank contributes 60.7% of the total package amount. And this number is 61.3% at the facility level. The median lead bank contribution at the package level (63.42%) is less than that at the facility level (70.1%). In our sample each package has 1.03 lead banks on average. The average number of all lenders (both lead and participant banks) of a package is about 6.20 and this number is very similar at the facility level (6.30). For participant banks, the average contribution is 8.8% (8.6%) per bank at the package (facility) level. Packages in our sample have an average total amount of $ million (in 2012 dollars). This implies that the average lead bank contribution at the package level is about $ million ( ). The median lead bank contribution is $66.56 million ( ). Other non-pricing loan characteristics display normal patterns. Loan covenants are structured at the package level. On average, each package in our sample contains two financial covenants and 1.66 non-financial covenants. 6 Performance pricing schedule and loan collateral requirements are negotiated at the facility level. A little more than half (53.6% and 54.9%) of facilities in our sample have a performance pricing schedule and are secured. At the facility level, about 64.4% of the facilities are leveraged. A loan is leveraged if it is documented by DealScan as a Highly Leveraged, Leveraged, or Non-Investment Grade loan. 81.7% of the facilities are syndicated in our sample. Here we implicitly assume that a lead bank can choose whether or not to structure a deal as a sole lender loan or a syndicated loan. Our main results are not sensitive to the inclusion or exclusion of the sole 6 See the definition of financial and non-financial covenants in Appendix A2. 13

15 lender loans. A loan package will be regarded as leveraged (syndicated) if all facilities in the package are leveraged (syndicated). Panel B of Table II reports the descriptive statistics for bank characteristics. The key variable of interest is the Total Capital Ratio, which is defined as total capital over riskweighted assets. It has a mean of 11.8% and a median of 11.2% for lead banks. The mean and median of Total Capital Ratio, 12.3% and 11.3%, are slightly larger for non-lead (participant) banks. For lead banks, the 25 percentile is 10.8% and the minimum value (not shown) is 9.1%, which indicates that lead banks are far from hitting the minimum capital requirements, which is 8%. Since our sample period includes the recent financial crisis, the high capital ratios of these banks may be partially due to capital accumulation during this abnormal period. It is also possible that banks with higher capital ratios choose to lend in the syndicated loan market or the banks choose to lend when their capital ratios are high. As alternative capital measures, the average Tier 1 Capital Ratio for lead banks is 8.8%. The average Leverage Ratio is 6.9%. Leverage ratios are smaller than the tier 1 capital ratios because the denominator is bank total assets without weighting by risk. The mean (median) value of lead bank total assets in our sample is $ ($217.19) billion. For participant banks, the mean (median) value is $ ($80.59) billion. Bank assets are much larger for our sample than the average size of commercial banks in the whole Call Report universe because larger banks are more active in the syndicated loan market. 7 The summary statistics of borrower characteristics, which are reported in Panel C of Table II, suggest that firms that borrow in the syndicated loan market are larger, more profitable, and more likely to have S&P ratings than average Compustat firms. These firms also have fewer growth opportunities. 7 In our sample, our averageloan size is over$400 million. A lead bank s averagecontribution to a package is about 0.06% of the bank s total asset and about 0.08% of a bank s total risk weighted assets. 14

16 II. The Baseline Results: The Effect of Bank Capital on Lead Bank Share In this section, we present the baseline results using lead banks only. Specifically, we regress the natural logarithms of Lead Bank Share (in percentages) on bank capital ratios measured at the calendar quarter that ends right before a loan s origination date along with other control variables. We focus on lead banks in our cross-loan analysis because it is the lead bank(s) that originates and arranges the loan. We include participant banks when we examine and compare the impact of capital on loan contributions within a loan package/facility. We estimate variations of the following model: Log(LeadBankShare) ijkt = α+β 1 BankCapital jt 1 +γ 1 X it 1 +γ 2 Y jt 1 +γ 3 Z ijkt +Bank,Year,Industry,LoanPurpose,and/orState YearFixedEffects+ǫ ijkt (1) where subscript i indexes the borrowing firm, subscript j indexes the lead bank, subscript k indexes the loan package/facility, and subscript t indexes time. The key variable of interest is Bank Capital, which is Total Capital Ratio in the baseline results. (We use Tier 1 Capital Ratio and Leverage Ratio in the robustness tests.) If bank capital has a positive effect on lending, we would expect its coefficient to be positive and statistically significant. X is a vector of borrower characteristics, Y is a vector of lead bank characteristics other than Bank Capital, and Z is a vector of loan characteristics. We report the results at the package level in Table III. In all regressions in Table III, we include loan origination year dummies to capture changes in the macroeconomic environment of bank credit demand and supply. We also include industry dummies defined according to the 2-digit SIC codes to control for industry specific effects on lead bank allocations. Package purpose dummies are also included to account for the possibility that banks with higher capital may prefer to involve in some loans with specific purposes. We cluster the standard errors at the lead bank level to account for the correlation between multiple loans made by 15

17 the same lead bank. In Column (1) of Table III under the OLS model, the estimated coefficient on Total Capital Ratio is positive and statistically significant at the one percent level. This suggests that, after controlling for other factors, a lead bank with more capital contributes more to a loan package. Economically, one standard deviation increase of Total Capital Ratio (0.025 or 2.5%, see Table II for the summary statistics) is associated with an increase of the lead bank holding by %=3.15%. This is equivalent to a 3.2% increase in lead bank share after adjusting for logarithm. This is equivalent to about 8% (3.2%/40%) of the standard deviation of the Lead Bank Share per package in our sample. Or in dollar value, one standard deviation increase in a lead bank s Total Capital Ratio is associated with an increase of $ million 3.2%=$13.16 million contribution to a package on average. Considering that a bank typically makes many loans in a year, the overall effect is economically significant if the lead bank s average contribution to each loan increases by about $13 million after one standard deviation increase of its Total Capital Ratio. In fact, given the effect of capital on lead bank share found here, it is reasonable to expect that bank capital can affect bank lending at the extensive margin, i.e., banks with lower capital levels may choose not to lend at all or not to lend to some borrowers. Therefore, the implied total effect of bank capital on lending may be even larger. In Column (2), we add the number of financial and non-financial covenants as additional controls for loan characteristics. The results are similar to those in Column (1). Although our framework is able to mitigate the concern that bank capital may be correlated with demand side factors, the OLS regressions do not address the potential problem that bank capital is correlated with unobserved bank characteristics. As a first step to address this problem, we include bank fixed effects in the regressions to control for the correlation between bank capital and unobserved time invariant bank characteristics. The results with bank fixed effects are presented in Columns (3) and (4) of Table III. In both columns, the coefficient estimates on Total Capital Ratio remain positive and have the same levels of statistical significance of one percent as their OLS counterparties. The magnitudes of the 16

18 effect of Total Capital Ratio in fact increase relative to the OLS estimates. To further alleviate the concern that demand side factors may simultaneously drive bank capital and the lead share, we include instead State Year fixed effects. The State Year fixed effects can absorb any confounding time-variant state level economic conditions that can simultaneously affect bank capital and loan demand and are otherwise unobservable. The results are presented in Columns (5) and (6) of Table III. In both columns, the coefficient estimates on Total Capital Ratio remain positive and statistically significant, which suggests that the positive effects of bank capital on lead share are unlikely to be driven by local economic conditions. In general, most of our control variables in the regressions carry expected signs. The coefficient estimates of Log (Bank Total Assets) are positive and statistically significant at the one percent level in Columns (1), (2), (5), and (6), which is consistent with the intuition that larger and more diversified banks are more capable of taking more risk. However, the effect disappears with bank fixed effects in Columns (3) and (4), which is probably due to the fact that bank size is highly autocorrelated. The coefficient of Bank Liquidity is also positive and statistically significant in all models, which is consistent with the argument that banks with more liquid assets are more likely to be able to fund loans on the margin. Bank ROA is positive in all columns and statistically significant in Columns (1), (2), (5), and (6), which is consistent with the notion that a more profitable bank has greater capacity to absorb negative capital shocks. The coefficient on Loan Charge-Offs is negative in all columns, which is consistent with the notion that greater losses resulted from previous lending is likely to curb a bank s incentive to lend in the future. The coefficients on Risk Weighted Assets are positive and significant at the one percent level in all columns. On the borrower side, the statistically significant negative coefficients on Rated Dummy show that a lead bank contributes less to a loan made to a rated firm, which is consistent with Sufi (2007) that a firms with credit rating has less information asymmetry so that the lead bank does not have to retain a large fraction. The coefficient on Cash Holdings is positive and statistically significant in all models, which is consistent with the notion that higher cash reserve could 17

19 exacerbate the free cash flow problem (Jensen (1986)) so that the lead bank holds more of a loan to facilitate monitoring. In Table IV, we show similar results at the facility level. Taken together, the baseline results in Tables 3 and 4 show that a bank with greater capital level is more likely to provide more credit. These results are less likely to be subject to the common criticism in the literature on bank lending channel that the link between bank capital and lending can be driven by demand side factors. First, our matched sample allows us to explicitly control for borrower characteristics that may matter for loan demand, which is usually not possible in studies based only on aggregate, regional, or bank balance sheet information(e.g., Bernanke et al. (1991), Berger and Udell (1994), and Peek and Rosengren (1997)). Second, the documented effect of bank capital on lending in this paper is already conditioned on the total amount of a package/facility, so the demand side factor should matter less for the lead bank s allocation in a loan. Lastly, one susceptible channel through which the demand side factors can work is that low quality firms may choose to borrow from a low capital bank since a low capital bank may lack the ability to effectively monitor the low quality firm. Such possibility can therefore lead to a positive relationship between bank capital and lending. However, the existing literature (e.g., Sufi (2007) and Ivashina (2009), among others) has shown that lead banks tend to contribute more when the borrower is riskier and/or when the asymmetric information problem is more acute. Thus, the selection issue is unlikely to explain the positive relationship between bank capital and lead bank share found here. Nevertheless, to further mitigate the concerns that the results are driven by the correlation between local economic conditions, which can affect loan demand and bank capital simultaneously, we conduct a falsification test with matched borrowers. For each loan in the sample, whenever possible, we match it to a loan that is made to another borrower who is in the same state, in the same 2-digit SIC industry in the same year and with the closest total book asset value, but borrows from a different lender. We then regress lead bank share of the matched loan on capital of the lead banks of the original loans. For each original loan in the sample, because the matched borrower is in the same state as the true borrower, the 18

20 capital ratios of the true lead bank should still have a positive effect on lead bank share of the matched borrower, if the baseline results were driven by the correlations between local economic conditions and bank capital. The falsification tests are presented in Table V. In this table, all bank characteristics are for the sample lead bank in the original sample, but the lead bank share and borrower characteristics are for the matched borrower. To save space, only coefficients and standard errors on bank characteristics are reported. Columns (1) and (2) present the regression results at the package level, and Columns (3) and (4) present the results at the facility level. The results show that: (1) The coefficients on Total Capital Ratio are statistically and economically insignificant; (2) None of the coefficients on bank characteristics are statistically significant. The results suggest that our baseline results are unlikely to be driven by the correlations between local economic conditions and bank capital. III. Alternative Identification Strategies To further establish causality, we use two alternative identification strategies in this section. We first use a within-package/within-facility estimation to further difference out any confounding loan and borrower characteristics that are otherwise unobservable. In the second identification method, we exploit plausibly exogenous variations in bank capital generated by the Troubled Asset Relief Program (TARP) and combine it with the within-package/withinfacility estimation to further establish the causality. A. Within-Loan Estimations Although we argue that the positive relationship between bank capital and lead bank share is unlikely to be explained by the demand side factors, it is still possible that some unobservable borrower characteristics, especially those related to asymmetric information about borrower quality, may be correlated with lead bank capital. This could cause the omitted variables problem. For example, it is possible that banks with higher levels of 19

21 capital are more willing to lend to firms with greater degree of information asymmetry, and at the same time, lead banks take more allocations to mitigate the moral hazard problem associated with monitoring and screening (Sufi (2007) and Ivashina (2009)). We address this omitted variables problem by looking into the within package/facility allocation differences between banks participating in the same loan package/facility. Empirically, we execute this strategy by estimating the following model: Log(BankShare) ijkt = α k +β 1 BankCapital jt 1 +γ 2 Y jt 1 +ǫ ijkt, (2) where subscript k indexes packages/facilities, and α k is the package/facility fixed effect. Y is a vector of bank characteristics. Note that once package/facility fixed effects are included, both borrower characteristics and loan characteristics drop out. By including package or facility fixed effects in the model specification, we thus remove any effect due to confounding borrower characteristics that are otherwise unobservable. The within-loan estimation is also free from any confounding effects caused by endogenously determined loan characteristics. 8 Therefore, any remaining differences in the relative loan shares between different lenders within a package/facility are unlikely to be driven by firm-level or loan-level factors. Rather, this difference is likely to be a function of bank-level factors. At the package level, we include 2,044 syndicated packages that have more than one lender per package between 1996 and 2012 because we require within package differences between multiple banks. 9 For the facility level analysis, we use 2,606 syndicated facilities with more than one lender per facility. We present the within-loan estimation results at the package level in Panel A of Table VI. In Column (1), we first include all banks (both lead and non-lead), and estimate the model with a lead bank dummy, which equals one if a bank is the lead bank in the package, and 8 For papers that use a similar approach for other estimation purposes, see, e.g., Ivashina and Sun (2011) and Lim et al. (2014) 9 According to DealScan, there are 3,516 syndicated packages (out of 4356) in our sample. But 1472 of them have only one lender recorded by DealScan. So for this analysis, we keep the 2044 ( ) syndicated packages with at least two lenders. 20

22 zero otherwise. In Column (2), we include only non-lead banks. In this sample, all non-lead banks are homogenous interms of their roles inapackage so we can ruleout any confounding effect resulted from unobservable difference between lead and non-lead banks during the loan arrangement process. 10 The coefficients on Total Capital Ratio in both columns are positive and statistically significant at the one percent level. This result suggests that, within the same loan package, a bank with a higher capital level contributes more to the entire package. Notethatthecoefficient ontheleadbankdummyincolumn(1)ispositive andstatistically significant, which suggests that lead banks tend to contribute more than participant banks. As pointed out by Sufi(2007), a lead bank s holding would signal its commitment to screening and monitoring of the borrower. In Columns (3) and (4), we further add bank fixed effects in the models. The coefficients on Total Capital Ratio are still positive but only marginally significant. Since the magnitudes of the coefficient estimates do not drop, the decrease in the significance levels is likely due to the decrease in statistical power. We re-estimate models with the facility level data, and the results are presented in Panel B of Table VI. The coefficients on Total Capital Ratio are all positive and statistically significant at least at the five percent level. The results are consistent with those shown in Panel A of Table VI. In summary, the evidence from the within-loan estimations in Table VI indicates that the effect of bank capital on lending is unlikely to be driven by unobserved borrower characteristics, or generally other demand side factors. B. Using TARP as a Quasi-Natural Experiment Although the baseline results and the within-loan estimation results can effectively separate the effects of bank capital from confounding demand side factors, they do not address the endogeneity concerns induced by other problems. In this subsection, we provide additional evidence on the casual effect of bank capital on lending by exploiting plausibly exogenous variations in bank capital levels generated by the Troubled Asset Relief Program (TARP). 10 We have only 59 packages with multiple lead banks so it is not statistically reliable to make inference from a similar test with lead banks only. 21