Research Design Sample Selection and Research Design 1. Sample Selection and Data

9 example, in July 2005, Burger King Corp. refinanced its credit facilities in one loan package, which includes a revolving facility, term loan A, and term loan B. We treat each facility as one observation given that each facility has its own specific loan characteristics. 3 We exclude financial-service firms given the uniqueness of their asset structure and revenue-generation process. We then merge the DealScan data with the Compustat database to obtain financial-accounting information, and with the interest-rate database on the Federal Reserve’s website to estimate the overall lending environment and market conditions. 4 Observations with missing values are excluded. This process results in a final sample of 15,346 bank loan facilities.

3.2. Research Design

To test our hypothesis about the impact of managerial ability on the cost of debt, we estimate the following model: Cost of Debt = β + β 1 Managerial Ability +  β i Loan Characteristics +  β j Borrower Characteristics +  β k Market Conditions + Industry FE + Year FE+ 1 Cost of Debt refers to the logarithm form of all-in-drawn spread over LIBOR. The all-in- drawn spread is defined as the annual spread paid over LIBOR for each dollar drawn from the loan. To increase the robustness of our results, we also use Total-Cost-of-Borrowing Berg, 3 This treatment may put more weight on borrowers who have one borrowing package containing multiple types of facilities. In a robustness test, we also partition samples by facility type and inferences remain intact. 4 http:www.federalreserve.govreleasesh15data.htm 10 Saunders and Steffen 2015 and credit rating as alternative proxies for cost of debt. Our main proxy for managerial ability is MA_DEA, a measure using Data Development Analysis DEA consistent with Demerjian, Lev, and McVay 2012. To construct MA_DEA, we first employ DEA to estimate firm efficiency within industries by comparing the sales generated by each firm conditional on the inputs used by the firm. Following Demerjian et al. 2012, we use sales as our output and seven other accounting inputs, consisting of cost of goods sold, SGA, PPE, operating leases, RD, goodwill, and other intangibles. Specifically, we solve the following optimization problem by industry year: max � = � � + � + + � + + � ���+ ℎ � ��� 2 We then separate managerial ability from firm efficiency estimated above by excluding the effects of key firm specific characteristics that may affect management’s efforts. Following Demerjian et al. 2012, we estimate the following regression model by industry: Firm Efficiency = β + β 1 lnTotal assets + β 2 Market Share + β 3 Free Cash-Flow I ndicator + β 4 lnAge + β 5 Business-Segment Concentration + β 6 Foreign-Operation Indicator + Year Fixed Effects + 3 Market share is calculated as firm’s revenue divided by aggregate revenue of all firms in the same Fama-French 48 industry. Free Cash-Flow Indicator equals one if the firm-year observation has positive free cash flow, zero otherwise. Age is one plus the number of years since IPO, which is set to be zero for private firms. Business-Segment Concentration is the squares of individual business segment sales to total sales, summed across all business segments. Foreign- 11 Operation Indicator equals one if the firm-year observation has foreign currency adjustments in its financial reports, zero otherwise. The residual from the estimation of equation 3 is then attributed to the management team. We then rank the original measure by deciles. A value of 0 represents the lowest decile of managerial ability and a value of 9 represents the highest decile of managerial ability. Following prior research e.g., Strahan 1999; Bharath, Sunder, and Sunder 2008; Francis, Hasan, Koetter, and Wu 2012, we control for a multitude of factors across three categories, namely loan characteristics, firm characteristics, and market conditions. In addition, we control for year and industry fixed effects. The Appendix provides detailed definitions for all our variables. Finally, standard errors are clustered by firm. 4. Empirical Results 4.1. Descriptive Statistics