Policy Uncertainty and Corporate Credit Spreads

Transcription

1 Policy Uncertainty and Corporate Credit Spreads Mahsa S. Kaviani a Lawrence Kryzanowski b Hosein Maleki c Pavel Savor d Abstract We find a significant positive relation between policy uncertainty and credit spreads. Macroeconomic conditions, including general uncertainty, do not explain this result, which also holds when we use instrumental variables to address endogeneity issues. The impact of policy uncertainty is greater for firms that are more exposed to government policies, such as those operating in regulation-intensive industries, paying high effective tax rates, and dependent on government spending. It is also stronger for firms that engage in political activities, and for those that are more dependent on external financing. We conclude that policy uncertainty has a significant effect on firms borrowing costs, with exposure to government policies representing an important channel. Keywords: credit spreads, policy uncertainty, regulation JEL: G12, G38 a Temple University, Fox School of Business, kaviani@temple.edu b Concordia University, John Molson School of Business, lawrence.kryzanowski@concordia.ca c Temple University, Fox School of Business, hmaleki@temple.edu d Temple University, Fox School of Business, pavel.savor@temple.edu We would like to thank Ronald Anderson, Pat Akey, Patrick Augustin, Elyas Eliasiani, Jan Ericsson, Huseyin Gulen, Kose John, Dong Hyun Kim, Stefan Lewellen, Babak Lotfaliei, Miriam Marra, David Newton and Baolin Wang for their valuable comments. Financial support from Senior Concordia University Research Chair in Finance, IFM2, and SSHRC is gratefully acknowledged. 1

2 1. Introduction Government policies have wide-ranging impact on firms and markets. They set the rules for how firms operate and compete, determine how firms are taxed and subsidized, and influence general macroeconomic conditions. Thus, when the level of disagreement among policy makers on fiscal, monetary, or regulatory issues increases, as it did in the U.S. and globally in the aftermath of the financial crisis, the overall level of uncertainty in an economy also rises. 1 How this policy-related uncertainty affects economic and financial activities is an important question that has attracted much interest among scholars, policy makers, and the general public. This paper investigates the effect of policy uncertainty on corporate credit spreads. We focus on credit spreads for two reasons. First, they are a direct measure of credit risk, and therefore allow us to estimate how and why policy uncertainty affects borrowing costs. Second, credit spreads also provide an opportunity to study the relation between policy uncertainty and economic outcomes. As argued by Faust et al. (2013) and Gilchrist et al. (2014), there is robust empirical evidence that credit spreads represent the most informative and reliable class of financial indicators of future economic activity. 2 Our goal is to determine whether uncertainty about government policies has a causal impact on credit spreads, and then to explore potential channels through which such uncertainty has an effect on credit risk. We thus study both the relation between policy uncertainty and credit spreads over time, as well as how the relation differs across firms. Policy uncertainty is not directly observable and is hard to distinguish from general uncertainty about economic conditions. One common proxy in the literature for periods of elevated policy uncertainty is elections. 3 A shortcoming of using election dates is their low 1 Baker et al. (2016) document a marked increase in the frequency of the word uncertainty as it relates to policy since the global financial crisis. 2 Adverse shocks to credit spreads predict declines in economic activity (Gertler and Lown (1999); Gilchrist et al. (2009); Gilchrist and Zakrajšek (2012); and Faust et al. (2013)). 3 Prior studies examine how national elections impact stock return volatility across countries (Boutchkova et al. (2012)); how elections affect investment (Julio and Yook (2012)); how political uncertainty in election years changes corporate investment sensitivity to stock prices (Durnev (2010)); and how uncertainty about 2

3 frequency, which means that they fail to effectively account for variation in policy uncertainty between election dates. Baker et al. (2016) (BBD thereafter) develop a policy uncertainty index (PUI ), which they construct from four components. The first component is a comprehensive measure of uncertainty based on the number of articles about economic policy uncertainty in ten large newspapers. The second component measures taxation uncertainty based on data from the Congressional Budget Office on expiring tax provisions, and the third and fourth components measure inflation and government expenditure uncertainty based on the dispersion in professional analyst forecasts. We use this PUI as our principal measure of policy uncertainty. BBD show that their index captures periods of heightened uncertainty about economic policies, such as elections, policy debates, or government policy changes. The index is used both by practitioners and in a number of recent academic studies on policy uncertainty (Pastor and Veronesi (2013); Gulen and Ion (2016); and Francis et al. (2014)). We begin by showing that the relation between policy uncertainty and credit spreads is strongly positive. A one-standard-deviation increase in PUI is associated with an increase in credit spreads of 62.8 basis point (bps), with a t-statistic (double-clustered by firm and time) of The R 2 for the univariate regression is 3%, indicating PUI has considerable explanatory power for variation in credit spreads. This result continues to hold when we include a variety of controls for bond-issue characteristics, firm characteristics, and financial conditions, as well as firm and year fixed effects. 4 In our most comprehensive specification, the standardized coefficient on PUI is 32.6 (t-statistic = 14.9), representing an economically and statistically important relation. 5 Comparing the impact of policy uncertainty on investment- and speculative-grade bonds, we find the effect exists for both types but is more pronounced for speculative-grade issues. gubernatorial elections affects the yields on municipal bonds (Gao and Qi (2012)). 4 To test the validity of the PUI, we compare its effect on credit spreads with that of other proxies for policy uncertainty suggested in the literature (including monetary, fiscal, and government policy uncertainty), and find that it is robust to any such controls. 5 The relation between PUI and credit spreads is equally strong when we focus on changes in these variables rather than levels in our regressions. 3

4 A one-standard-deviation increase in PUI increases speculative-grade credit spreads by 50.5 bps vs bps for investment-grade issues. We also decompose the overall PUI into its four constituent components, and find that the positive relation between policy uncertainty and credit spreads exists for all four. Our findings so far document a positive relation between policy uncertainty and credit spreads, but do not address the issue of causality. One potential explanation for our results is omitted variables about economic conditions or economic uncertainty. Both policy uncertainty and credit spreads increase during downturns and during periods of elevated economic uncertainty, so the two could be positively correlated even in the absence of any direct relation between the two. However, our main results hold (even though coefficient estimates are lower) when we include controls for macroeconomic conditions, suggesting they do not explain the relation between PUI and spreads. Similarly, the relation between PUI and spreads remains highly significant when we add controls for economic uncertainty, such as implied S&P 500 return volatility, GDP growth forecast dispersion, idiosyncratic volatility, and an elections indicator. While encouraging, the above tests do not conclusively rule out general economic uncertainty, as distinct from policy uncertainty, as an explanation for our findings. It is possible that the economic uncertainty proxies we use do not fully capture other sources of economic uncertainty. To further address endogeneity concerns about the causal impact of policy uncertainty on credit spreads, we conduct two tests, motivated by Gulen and Ion (2016). First, we note that economic conditions in the U.S. are highly correlated with those in Canada, while policy uncertainty in the two countries is likely less correlated. We exploit this opportunity by regressing the U.S. PUI on the Canadian PUI, and then use the regression residuals as an alternative index of policy uncertainty, which ideally does not reflect underlying general economic uncertainty. Second, we adopt an instrumental variables approach. We identify three instruments for U.S. policy uncertainty: the relative power of the two main U.S. political parties, the level of political polarization in the country, and the interaction 4

5 of these two indices. The relation between PUI and spreads remains strongly positive with both of these approaches, alleviating concerns that endogeneity explains this finding. We next study possible mechanisms through which policy uncertainty affects credit spreads by exploring how the impact of policy uncertainty differs across firms. Broadly speaking, government policies can affect firms in two ways: directly, through their specific impact on a given firm, and indirectly, through their impact on the broader economic environment. We first focus on the former channel, which has so far received little attention in the literature. More specifically, we expect that policy uncertainty should matter more for those firms that are more exposed to government policies. For such firms, increases in uncertainty about government actions should increase uncertainty about the firms future prospects, which should translate into higher credit risk. We identify three types of firms that are especially exposed to government policies: firms in industries that are subject to more government regulations, firms paying high effective tax rates (ETRs), and firms that are dependent on government spending. The first category covers firms that are most affected by regulatory policies, the second those that are most sensitive to taxation policies, and the third those that are most exposed to spending policies. The relation between PUI and spreads is significantly stronger for these three categories of firms: the PUI coefficient is 29% higher for firms in regulation-intensive industries (with a t-statistic of 1.7 for the difference), 61% higher for high-etr firms (t-statistic = 4.2), and 41% higher for firms dependent on government spending (t-statistic = 2.1). These findings are consistent with the hypothesis that all three types of policy uncertainty influence firms borrowing costs. We also show that policy uncertainty has a greater impact on the credit spreads of politically active firms, such as those engaging in lobbying activities or making campaign contributions. For example, lobbying firms have a PUI coefficient that is 149% higher (t-statistic= 6.9) than that of non-lobbying firms. Since the intensity of firms political activities should be positively related to their exposure to government policies, this result provides further support for the conclusion that policy uncertainty is especially important 5

6 for spreads of firms exposed to these policies. In addition to its direct impact on firms, policy uncertainty can have an indirect effect on firms borrowing costs. If heightened uncertainty about economic policies makes it harder for firms to issue debt or equity, then firms dependent on external financing would abandon or delay profitable investments. This, in turn, would increase these firms credit risk and consequently credit spreads. We measure firms dependence on external financing as in Duchin et al. (2010), and find support for this hypothesis. The relation between policy uncertainty and spreads is 177% stronger for firms whose dependence on external financing is above that of the median firm (t-statistic = 5.8), representing an economically and statistically meaningful effect. Credit spreads reflect compensation for expected default losses and a risk premium. We decompose spreads into these two components, and find that policy uncertainty has a significant positive effect on both. These results indicate that increases in policy uncertainty are associated with a higher probability of default as well as a higher credit risk premium, and are thus consistent with the hypothesis that policy uncertainty affects credit spreads through both cash flows and discount rates. Our paper makes two important contributions to the existing literature. First, we contribute to the growing literature on the impact of policy and political uncertainty on corporate investment and financing decisions (Gulen and Ion (2016); Bonaime et al. (2016); Bradley et al. (2016); Jens (2016); Francis et al. (2014); Cao et al. (2013); Pastor and Veronesi (2013); Pastor and Veronesi (2012) Julio and Yook (2012); Durnev (2010); and Gao and Qi (2012)) by showing how corporate credit spreads and cost of capital react to changes in policy uncertainty. Our results are consistent with frequent speculation that the recent weak recovery in the U.S. after the financial crisis is to some extent due to uncertainty over fiscal policies and regulatory reforms. 6 We also add to the related literature on the effects 6 See, for example, the Minutes of the Federal Open Market Committee, August 9, 2011, federalreserve.gov/monetarypolicy/fomcminutes htm; and Becker et al. (2010), Uncertainty and the slow recovery, in the Wall Street Journal. 6

7 of political connections on corporate decisions and performance (Akey and Lewellen (2016); Reza et al. (2016); Borisov et al. (2015); Brogaard et al. (2015); and Akey (2015)) by highlighting how higher exposure to different government policies amplifies the response of credit spreads to policy uncertainty changes. Second, we contribute to the literature on the determinants of credit spreads (Elkamhi et al. (2012b); Ericsson et al. (2009); Chen et al. (2007); Campbell and Taksler (2003); ; Collin-Dufresne et al. (2001)) by finding that policy uncertainty, a factor which previous papers did not study, has significant explanatory power for corporate credit spreads even after controlling for known credit spread determinants. Collin-Dufresne et al. (2001) and Huang and Huang (2012) show that variables proposed by structural models do not fully explain either levels of or changes in credit spreads, with Collin-Dufresne et al. (2001) conjecturing that a missing systematic factor accounts for a large fraction of variation in spreads. Our results suggest that policy uncertainty represents a potential candidate for this missing factor. This implication is consistent with Pastor and Veronesi (2012) and Pastor and Veronesi (2013), who develop models in which uncertainty about government policy and political uncertainty, both of which should be captured by the policy uncertainty index we use, increase equity risk premia. 7 The remainder of the paper proceeds as follows. Section 2 describes our data sources and sample construction. Section 3 reports the findings on the relation between policy uncertainty and credit spreads. Section 4 addresses endogeneity concerns. Section 5 explores crosssectional variation in the effect of PUI on spreads. Section 6 decomposes credit spreads into a component compensating investors for expected default losses and the bond risk premium. Section 7 investigates the robustness of our results, and Section 8 concludes. 7 Kelly et al. (2016) focus on elections and global summits, and provide evidence that political uncertainty is priced in the equity options market. 7

8 2. Data and Sample Construction 2.1. Data Sources Our primary source for bond data is the Trade Reporting and Compliance Engine (TRACE) database. This database is based on a regulatory requirement mandating increased transparency in the secondary bond market, and provides bond transaction data on a daily basis, including features such as transaction price, yield to maturity, and bond maturity date. Following Chen et al. (2010), we compute bond yields from bond prices and remove data points where our estimate differs from TRACE yields by more than 5%. We focus only on completed trades between 2002 and 2015, and therefore exclude from the sample trades marked as canceled, corrected, or suspended (similar to Edwards et al. (2007)). We also remove callable and convertible bonds, as in Campbell and Taksler (2003). If a bond is traded more than once in a given day, we compute its yield and price based on the average of transactions completed on that day. The data frequency in this paper is monthly, so our monthly measures of bond yields and prices are averages over each month of daily yields and bond prices. To include bond characteristics, we merge the TRACE database with the Fixed Income Securities Database (FISD) using the 9-digit issuance (CUSIP) codes that are common to both databases. This adds Moody s ratings and a variety of bond characteristics to our sample, such as coupons, maturities, and time of issuance. Next, we merge this database with the quarterly COMPUSTAT North America database. We retain in the merged TRACE- FISD sample all firms that have accounting information in COMPUSTAT. We remove bonds with maturities of less than one year, and winsorize the data at the 1% level to deal with possible outliers, similar to Campbell and Taksler (2003) and Schestag et al. (2016). Credit spreads are computed by subtracting the yield of the closest-maturity Treasury bond from the yield of each bond, where the former is obtained from the Federal Reserve Board website. 8

9 2.2. Measuring Policy Uncertainty We measure economic policy uncertainty using the aggregate index of BBD, who construct their policy uncertainty index from four components. The first component is a comprehensive measure of uncertainty based on an index of search results for articles on economic policy uncertainty in ten major newspapers. 8 This component has a weight of 0.5 in the overall PUI. The second component measures taxation uncertainty based on a list of expiring tax provisions compiled by the Congressional Budget Office. 9 The third component measures monetary policy uncertainty using as a proxy the Consumer Price Index (CPI ) forecast dispersion computed from individual forecasts in the Federal Reserve Bank of Philadelphia s Survey of Professional Forecasters. The fourth component measures federal/state/local expenditures forecast dispersion, again computed from individual forecasts in the Federal Reserve Bank of Philadelphia s Survey of Professional Forecasters. The last three components each carry a weight of 1/6 in the overall PUI. More details on the PUI construction are available at BBD show that their PUI captures effectively changes in policy uncertainty not only around elections, but also during other periods of elevated policy uncertainty, such as the debates over the debt ceiling, Gulf wars, the Federal Reserve s tapering of quantitative easing, and the U.S. government shut-down. BBD also perform a series of tests and validity checks to show that PUI appropriately proxies for economic policy uncertainty. In the first set of tests, they change the search term uncertainty to equity price, stock market, or stock price, and show that the index created using this method has a correlation of more than 70% with the VIX index (a measure of implied volatility for the S&P 500 index). Second, BBD perform human audits of newspaper articles used to construct the index. They 8 BBD consider keywords such as uncertain or uncertainty that are in used in conjunction with terms like economic as well as policy-related terms such as Congress, deficit, or Federal Reserve. 9 Uncertainty about tax code expiration is an important source of uncertainty for both households and firms. Congress can decide to extend a temporary tax measure or let it expire, often deciding at the last moment. This introduces significant uncertainty about future taxes that is captured by this component of PUI. 9

10 find that only in 1.8% of cases the human inference on the direction of policy uncertainty change is different from that of the mechanically constructed index. Third, BBD examine the validity of choosing newspapers as the main source of information. Particularly, they test whether the reflection of policy-related news in newspapers is influenced by the political positions of newspapers at the times when the ruling party has different political ideologies than a given newspaper, and find no evidence for this hypothesis. Finally, BBD change their primary data sources and re-create the index as a test of robustness. For example, using the Fed s Beige Books as an alternative source of data, they find that the correlation of the new index with the original one is as high as 80%. 10 We also test the PUI in a variety of ways, including comparing the index with other monetary-, fiscal-, or government-policy uncertainty proxies introduced in the literature. Overall, our results confirm the BBD conclusion that the PUI is an appropriate and robust indicator of economic policy uncertainty Summary Statistics Table 1 reports the summary statistics for the main variables used in our analysis. Panel A shows the summary statistics for bond- and firm-level characteristics in three sections. The first section covers all bonds in our sample, while the second and third sections cover investment-grade and speculative-grade bonds, respectively. As expected, speculative-grade bonds have higher credit spreads and coupons, shorter maturities, and are issued by firms with higher leverage and idiosyncratic risk. [Insert Table 1 About Here] Panel B provides the summary statistics for macroeconomic variables and proxies for policy uncertainty, all of which exhibit considerable variation over our sample period. Panel 10 The Beige Books are published by the Federal Reserve. Each District Fed collects anecdotal evidence on the state of the economy through a report. These reports are generated using interviews with bank and branch directors, economists, and other experts. This information is then summarized by the District Fed in the Beige Book. The books are published eight times per year. 10

11 C reports the summary statistics for bond yields, term spreads, and maturity-matched credit spreads (total yield minus the closest-maturity Treasury bond yield) separately for low- and high-policy uncertainty regimes (based on whether the PUI in a given month is below or above the median PUI ). All three variables are significantly higher in high-pui periods. Of special interest are credit spreads, which are on average 1.29% higher (the difference in medians is 0.82%), suggesting there exists a strong positive relation between policy uncertainty and credit spreads. We explore this hypothesis in more detail in the next section. 3. Impact of Policy Uncertainty on Corporate Credit Spreads In this section, we study the relation between economic policy uncertainty and corporate credit spreads. Given the large set of potential determinants of credit spreads, we include in our analysis various bond-level, firm-level, and macroeconomic control variables, typically selected from previous work, including Pogue and Soldofsky (1969), Pinches and Mingo (1973), Leland (1998), Blume et al. (1998), Campbell and Taksler (2003), Collin-Dufresne et al. (2001), Chen et al. (2007), and Ericsson et al. (2009). Firm-level controls are market leverage, operating income-to-sales ratio, total debt ratio, and pre-tax interest coverage dummies. These control variables are used widely in the literature on credit spreads. High pre-tax interest coverage and operating income to sales indicate that the firm is in better financial health and consequently should enjoy lower credit spreads. Higher leverage implies more financial risk and should be associated with higher credit spreads. Since the highest frequency for accounting data is quarterly, we compute a monthly market leverage ratio by dividing total debt (short- plus long-term debt, as of quarter-end) by average firm value, where the equity portion is updated daily (as in Ericsson et al. (2009)). Instead of using a continuous variable, we construct four interest coverage dummies. The reason for this approach, as argued by Blume et al. (1998) and Campbell and Taksler (2003), 11

12 is that changes in interest coverage have non-linear effects on credit spreads. For example, an increase in interest coverage from four (BBB-rated bonds) to six (A-rated bonds) can result in a bond upgrade, while a similar change from 20 to 22 has almost no effect on a bond s rating. Low pre-tax coverage values can therefore be much more informative about the issuer s risk. The interest coverage dummies are defined for values less than five, between five and ten, between ten and 20, and above 20. We also control for a firm s stock return in a given month as well as firm idiosyncratic volatility, measured as the standard deviation of daily excess returns relative to the CRSP value-weighted index over 180 days prior to the bond transaction date. Campbell and Taksler (2003) show the standard deviation of daily excess returns has a positive impact on credit spreads. We use three main macroeconomic controls: the yield of the Treasury bond that is closest in maturity to that of a given bond issue, the term slope, and the S&P 500 index return. The term slope is computed as the difference between ten- and two-year maturity Treasury yields. We expect a negative relation between the level of Treasury rates and credit spreads, as discussed by Longstaff and Schwartz (1995), since higher interest rates raise the riskneutral drift of the firm value and thus reduce the risk-neutral default probabilities. This should naturally be translated into lower credit spreads. The relation between the term slope and credit spreads is less clear, as the slope of the term structure can be a measure of economic uncertainty, resulting in a positive impact, as well as an expectation of future short-term rates, which would have a negative impact (Collin-Dufresne et al. (2001)). We include S&P 500 returns to account for general market conditions. We expect a negative relation between market returns and bond yields, since higher returns indicate better economic conditions and therefore better firm performance, which in turn should push credit spreads down. We control for a bond s credit rating, as it directly affects credit spreads. Ederington et al. (1987) and other subsequent papers find that credit ratings help explain cross-sectional 12

13 differences in credit spreads even after controlling for firm and issue characteristics, with a clear negative relation between changes in a bond s rating and changes in its credit spread. We use two credit rating measures: the annual S&P Domestic Long Term Issuer Credit Rating from COMPUSTAT, and Moody s bond rating at issuance from the FISD database. We convert alphabetical ratings to numerical equivalents as in Lerner et al. (2011). To do so, we assign numbers to bond ratings so that each notch difference in alphabetical ratings translates into one unit change in the numerical measure. With this method, for example, Aaa equals one, Aa1 equals two, and C equals 21. As higher value indicates a lower credit rating, we predict a positive relation between this numerical equivalent and credit spreads. We also control for a bond s liquidity, though there is no clear theoretical prediction about its relation to spreads. There exist a variety of methods in the literature for measuring bond liquidity. For example, Campbell and Taksler (2003) use issue size to proxy for liquidity. Bond age is another potential proxy (Mansi et al. (2012); Sarig and Warga (1989); and Adams and Mansi (2009)), even though there is no straightforward relation between a bond s age and its liquidity. Guntay and Hackbarth (2010) use the number of months a bond is traded prior to the bond s transaction date. Addressing the lack of consensus in measuring bond liquidity, Schestag et al. (2016) compare common liquidity measures in the literature and find that the best are those in Corwin and Schultz (2012), Roll (1984), and Hasbrouck (2009). Accordingly, we use the proxy in Schestag et al. (2016), which is based on the ratio of daily high to daily low prices, as our primary liquidity measure. 11 We include coupon rates since bonds with higher coupon rates are taxed more during their life (Elton et al. (2001) and Campbell and Taksler (2003)). This leads to lower desirability of these bonds compared to bonds with lower coupon rates. We further control for other general bond characteristics, including maturity and the putability feature. Depending on the term spread, bond maturity can impact its yield. Putable bonds can represent less risky investments because the holder has the option to sell the bond back to the issuer before the 11 Our results are the same with alternative liquidity measures. 13

14 bond matures (of course, the inclusion of this feature is likely influenced by a firm s risk characteristics). Finally, we take into consideration election periods by including an election dummy, which is the traditional indicator of policy uncertainty in the literature (see, for example, Boutchkova et al. (2012) and Julio and Yook (2012)). This dummy variable equals one during the four months before and including the month of presidential and congressional elections, and is zero otherwise. We provide detailed descriptions of all the above variables in Appendix Full Sample Results We begin our analysis by estimating the relation between policy uncertainty and credit spreads. Equation 1 shows our primary regression specification. The independent variable of main interest is PUI, our proxy for the level of policy uncertainty. As described above, we control for various issue-level, firm-level, and macroeconomic determinants in most specifications. To account for spurious regression results, where the relation between PUI and credit spreads is driven by changes in other macro variables that simultaneously affect both spreads and policy uncertainty, we use lagged PUI values. Our main specification incorporates firm fixed effects, which should mitigate the possible unobserved effects of firm-level determinants that are not captured in related controls and also address concerns about persistence in firm-level controls (Lemmon et al. (2008)). We also include year fixed effects to alleviate worries that our findings are driven by common trends in PUI and spreads. Our base regression specification is: CREDIT SP READ i,t = β 0 + β 1 P UI t 1 + β 2 F IRMCT RL i,t + β 3 BONDCT RL i,t + β 4 MACROCT RL t + η i + m t + ɛ i,t. (1) P UI t 1 is the lagged policy uncertainty index. F IRMCT RL i,t are firm-level controls including market leverage, operating income to sales, total debt ratio, pre-tax interest coverage 14

15 dummy variables, monthly firm stock return, and firm idiosyncratic risk. BONDCT RL i,t are bond-level characteristics, which include its credit rating, liquidity, coupon rate, maturity, and the putability feature. Finally, MACROCT RL t includes macroeconomic variables such as the term slope, S&P 500 index return, the maturity-matched Treasury bond rate, and the election dummy variable. η i and m t capture the firm and year fixed effects, and ɛ i,t is the error term. We compute t-statistics using double-clustered standard errors by firm and year. Table 2 reports coefficient estimates and t-statistics for the regression specifications given by Equation 1. We present standardized coefficients to more clearly show how credit spreads change (in basis points) in response to a one-standard-deviation change in each of the explanatory variables. Column (1) shows the univariate regression results, using BBD s policy uncertainty index as the main explanatory variable. The PUI coefficient is 62.8 (t-statistic = 17.6), representing an economically meaningful and highly statistically significant effect. Adjusted R 2 is 3%, indicating that PUI has non-trivial explanatory power for variation in credit spreads. Column (2) adds firm and year fixed effects, which have little effect on the relation between PUI and spreads. The PUI coefficient remains significant both in terms of coefficient magnitude (42.4) and its t-statistic (17.7) when we include macroeconomic and bond-specific controls in Column (3). One of our controls is an election dummy, which is the traditional proxy for policy uncertainty. The estimated coefficient for this dummy is positive (12.4, with a t-statistic of 3.9), but its inclusion has only a marginal impact on the coefficient on PUI. Column (5) reports the findings for our most comprehensive regression specification, which includes all of our bond-, firm-, and macro-level control variables, as well as firm and year fixed effects. Crucially, policy uncertainty still has a large positive impact on credit spreads (32.6, with a t-statistic of 14.9). The results in Table 2 generally confirm the predictions and findings in the literature about how various bond-, firm-, and macro-level determinants affect credit spreads, indi- 15

16 cating that our sample is representative and comparable to those used in other studies. Deterioration in bond credit rating increases credit spreads, consistent with the intuitive expectation that lower-rated bonds have higher credit spreads. Higher contemporaneous firm and S&P 500 returns, as expected, reduce credit spreads, while higher idiosyncratic risk is associated with higher spreads. There exists a positive relation between bond maturity and firm leverage and credit spreads, in line with previous findings. Higher interest coverage ratios are associated with lower spreads. The relation between liquidity and credit spreads is positive, a potentially puzzling finding that may arise because we are controlling for contemporaneous liquidity and the direction of causation can flow either way. As suggested by Longstaff and Schwartz (1995), the yield of the maturity-matched Treasury bond has a negative influence on credit spreads. [Insert Table 2 About Here] 3.2. Investment- vs. Speculative-Grade Bonds We next investigate whether the relation between policy uncertainty and credit spreads depends on credit quality by estimating Equation 1 separately for investment- and speculativegrade bonds. Table 3 presents our findings. In Panel A, we use the S&P Domestic Long Term Issuer Credit Rating, and in Panel B we use Moody s bond rating at issuance. The investment-grade sample includes bonds with Moody s (S&P) credit rating of Baa3 (BBB-) and higher, and the speculative-grade sample includes those with credit rating of Ba1 (BB+) and lower. Our results show that the impact of policy uncertainty on credit ratings is greater for speculative-grade bonds. Even after controlling for all the bond-, firm-, and macro-level determinants of credit spreads, a one-standard-deviation increase in policy uncertainty is associated with a 50.5 bps increase in speculative-grade credit spreads compared to a 29.5 bps rise for the investment-grade sample. The higher sensitivity of spreads for speculativegrade bonds is not surprising, as these bonds are generally closer to default so that a larger 16

17 fraction of their credit spreads reflects compensation for firm-specific default probability. For example, by studying a set of structural models, Huang and Huang (2012) conclude that credit risk constitutes a small fraction of investment-grade bonds credit spreads, whereas it accounts for a much larger portion of spreads for speculative-grade bonds. [Insert Table 3 About Here] 4. Establishing Causality The results in the previous section document a strong positive association between policy uncertainty and credit spreads. In this section, we study the causal relation between these two variables, and specifically address whether changes in economic policy uncertainty cause changes in corporate credit spreads. To infer causality, we need to alleviate any possible endogeneity concerns. The principal such concern is that our results are driven by omitted variables, in particular unobserved economic uncertainty or business cycles effects, that affect the PUI and credit spreads simultaneously. For example, if uncertainty about general economic conditions increases both policy uncertainty and credit spreads, we can observe a positive relation between PUI and spreads, even in the absence of any causal relation between the two. Similarly, the PUI itself may reflect economic uncertainty that is not related to policy, which would introduce a measurement error bias. Another potential problem is simultaneity in the form of a reinforcing loop between PUI and credit spreads. Specifically, an increase in policy uncertainty can push spreads higher, and at the same time higher spreads may be interpreted by market analysts as indicators of higher policy uncertainty and therefore be reflected in the news. Thus, simultaneity is an eminent concern, since the main component of the PUI is a news-based index constructed by counting related articles in newspapers. 17

18 4.1. Omitted Variables Business Cycles It is well-documented that credit spreads are counter-cyclical (Gulen and Ion (2016); Baker et al. (2016); and Bloom (2014); ). At the same time, policy uncertainty can also escalate during economic downturns. Bad economic times create more incentives for governments to seek policy changes, and consequently can lead to higher policy uncertainty. To corroborate this idea, Bloom (2014) shows that the PUI is 51% higher during recessions. 12 Furthermore, Pastor and Veronesi (2012) conjecture that policy uncertainty is more likely to occur when the economy is in the downturn and as compensation investors demand higher risk premiums. This simultaneous rise in corporate credit spreads and policy uncertainty during economic downturns imposes a challenge for the causal interpretation of our results, as our regressions can just be capturing the effects of general economic conditions. The use of lagged values partially addresses these concerns, but our findings can still be influenced by omitted variables that persist over multiple periods (such as business cycles). Thus, we add to our regressions several macroeconomic variables that capture fluctuations in economic conditions: expected GDP growth, unemployment, and one-year inflation from the Survey of Professional Forecasters by the Federal Reserve Bank of Philadelphia, and the Michigan Consumer Confidence Index from the University of Michigan. We also keep all the controls used in Table 2, including firm and year fixed effects. We present the findings in Panel A of Table 4. Columns (1) through (4) include each of the new controls separately, and Column (5) includes all four simultaneously. The results confirm the hypothesis that spreads increase in economic downturns: they are negatively related to expected GDP growth and consumer confidence, and positively related to expected unemployment. Credit spreads are also high when expected inflation is low, as was the case 12 In fact, increases in uncertainty during recessions are not limited to policy uncertainty. Bloom (2014) finds that all different measures of uncertainty surge during recessions. Uncertainty increases even at the micro level (individual firms, plants and industries). Moreover, Campbell et al. (2001) document that individual stock return volatility increases more than 50% in recessions compared to booms. 18

19 during the Great Recession of Most importantly for our purposes, the PUI coefficient remains positive and strongly significant in all specifications, ranging from 6.4 to 29.7 bps. In the most comprehensive specification, the coefficient magnitude is however lower than before, suggesting that economic conditions explain a portion of the positive association between policy uncertainty and credit spreads. [Insert Table 4 About Here] Our sample covers the financial crisis of , a period during which both policy uncertainty and credit spreads spiked to high levels. To ensure our findings are not driven exclusively by this unique period, in Panel B we repeat our analysis from Table 2 but exclude years 2008 and The impact of the PUI is still large and significant at 12.9 bps (tstatistic = 6.7). Combined with those in Panel A, these results show that the relation between policy uncertainty and credit spreads is robust to the influence of general economic conditions Economic Uncertainty Another problem with the causal interpretation of the impact of PUI on credit spreads is the possibility that our results are capturing changes in general economic uncertainty rather than policy uncertainty per se. Policy uncertainty likely responds to shifts in general economic uncertainty, which also affects credit spreads. Thus, an alternative explanation for our findings on the relation between PUI and spreads is that both are driven by economic uncertainty. We begin addressing this potential problem by first confirming that our results hold when we control for a variety of economic uncertainty indicators. In the next sections, we then conduct further tests using instrumental variables and policy uncertainty in Canada. Our first proxy for economic uncertainty is the VXO index, which measures the implied volatility of the S&P 100 index. High levels of VXO should coincide with periods of heightened economic uncertainty. We use GDP growth forecast dispersion from the Survey of Professional Forecasters as an additional measure of economic uncertainty. Higher GDP 19

20 growth forecast dispersion indicates more uncertainty about economic conditions and thus can capture an additional aspect of economic uncertainty not reflected in the VXO. We show the results with these controls in Table 5. The first two columns include each of the variables separately and Column (3) adds all of them together. Both measures of economic uncertainty have a significant positive impact on credit spreads, both separately and together, corroborating the idea that increased economic uncertainty leads to higher spreads (and also validating our measures). The relation between the PUI and credit spreads stays positive and significant, showing that it is robust to the inclusion of economic uncertainty controls. The PUI coefficient magnitude (ranging from 23.3 to 27.1) is only slightly lower than its estimate without such controls (32.6). These findings support the hypothesis that economic uncertainty does not explain the relation between policy uncertainty and spreads. However, while the results are encouraging, it is still possible that we are simply not properly controlling for economic uncertainty, and therefore we further address the potential endogeneity problems below. [Insert Table 5 About Here] 4.2. Residual Policy Uncertainty One important limitation of BBD s policy uncertainty index is possible measurement error. Although this index is intended to measure economic policy uncertainty, it can potentially also capture other things. For example, the PUI may be contaminated by economic uncertainty or other factors that influence corporate credit spreads and are unrelated to policy uncertainty. While controlling for economic uncertainty in the previous section partially addresses this concern, in this section we introduce an additional test to account for remaining concerns about measurement error in the PUI, as suggested by Gulen and Ion (2016). The basic idea underlying this test is that the U.S. and Canada have closely related economies with high levels of trade between the two countries. The two countries have the 20

21 largest trade relationship in the world. 13 Thus, many of the economic shocks that affect one of these countries should also influence the other. Due to the much greater size of the U.S. economy, it is more likely that shocks to the U.S. economy impact the Canadian economy than the reverse. At the same time, policy uncertainty shocks are arguably more contained within a country s borders. This provides an opportunity to take out the portion of the PUI that is capturing economic uncertainty unrelated to policy. We do so by regressing the U.S. PUI on its Canadian counterpart, and then using the regression residuals as an alternative policy uncertainty index, controlling for a variety of macroeconomic variables. Our regression specification is: USP UI t = γ 0 + γ 1 CANP UI t + β 3 MACROCT RL t + m t + ɛ t. (2) USP UI t and CANP UI t are the BBD U.S. and Canada policy uncertainty indices, respectively. MACROCT RL t includes our macroeconomic controls: the three-month Treasury rate, term spread, expected GDP growth, and inflation. m t captures the year fixed effect. We use the residual ɛ t as the new policy uncertainty index and refer to it as the regression-based PUI (RPUI). This RPUI represents a cleaner version of the original PUI in terms of possible contamination with policy-unrelated components. We replace the PUI with RPUI in our main regression specification: CREDIT SP READS i,t = β 0 + β 1 RP UI i,t 1 + β 2 F IRMCT RL i,t + β 3 BONDCT RL i,t + β 4 MACROCT RL t + η i + m t + ɛ i,t. (3) As in our previous tests, we include firm and year fixed effects, and compute standard errors clustered at both firm and year levels. Table 6 presents the results, which show that RPUI has a significant impact on credit spreads. In the most comprehensive specification in Column (4), the coefficient estimate is positive and highly significant at 45.3 bps (t-statistic 13 See the Congressional Research Service Report to Congress in 2008 at: nationalaglawcenter.org/wpcontent/uploads/assets/crs/rl33087.pdf. 21

22 = 16.2). We conclude that measurement error does not explain our results on the relation between PUI and spreads. [Insert Table 6 About Here] 4.3. Relative Political Power and Polarization Our findings so far in this section alleviate concerns that endogeneity drives the observed relation between policy uncertainty and credit spreads, but do not fully resolve them. Our proxies for economic uncertainty may not capture all its sources, so it is still possible that some unobserved dimension of economic uncertainty determines both policy uncertainty and credit spreads. A similar logic applies to business cycle conditions. Policy uncertainty in Canada may be too dependent on political developments in the U.S., so that the residual PUI still reflects underlying general economic conditions. To address the above limitations and thus further address any remaining endogeneity concerns, we now introduce instrumental variables. We need to find variables that have a strong impact on policy uncertainty, but do not affect credit spreads through any other channel than their relation with policy uncertainty. Based on work in the political science literature, we identify three such variables. Our first instrumental variable is the relative legislative power of the two major U.S. political parties in each year. The data comes from the Swank (2013) database Comparative Political Parties Dataset, which provides the relative legislative strength of different political parties across 21 countries between 1950 and This database categorizes the ideological position of parties in every country into Right, Center, and Left. In the U.S., the Democratic and Republican parties are categorized as Center and Right leaning, respectively. Our key variable of interest is the difference in legislative power between the majority and the minority party (DIFFGS), which we compute as the difference in the percent of legislative seats that are occupied by each party. Given that coverage in the Comparative Political Parties Dataset ends in 2013, we update its data for the U.S. through

23 The idea here is that a reduction in the relative power of the majority party can increase uncertainty about policy expectations by making the resolution of sensitive issues and possible gridlocks more difficult (Erikson et al. (1989)). 14 At the same time, there is no obvious channel through which relative political power would directly affect credit spreads, making this measure a suitable candidate for an instrumental variable. Since we expect lower values of this index to lead to higher policy uncertainty, we multiply it by -1 for all our tests. As expected, this measure is highly positively correlated with the PUI. Our next instrumental variable is the degree of political polarization of the two main parties by Poole and Rosenthal (1985), which is first used for a similar purpose in Gulen and Ion (2016). The DW-NOMINATE variable in this database measures the different parties ideological positions over time based on their legislators voting patterns. 15 The political polarization measure has two dimensions. The first dimension, as Poole and Rosenthal (1985) argue, can be interpreted as the positions of legislators about government intervention in the economy, while the second dimension addresses the conflict between the North and South on slavery (before the Civil War) and civil rights for African Americans (from 1930s to mid-1970s). Recently, a dominant fraction of total polarization can be attributed to the first dimension (Lewis and Poole (2004); Carroll et al. (2013)), which is also more relevant for our tests that focus on economic policy uncertainty. Consequently, we focus on the first dimension (as in Gulen and Ion (2016)). We construct our instrument by subtracting the DW-NOMINATE score of the Democratic party from that of the Republican party. We expect that higher degree of political polarization complicates policy- and law-making, as the building of coalitions with their attendant compromises becomes more difficult (Mc- Carty (2004)). This in turn should lead to greater uncertainty about future policies. As with the first instrument, there is no clear channel through with political polarization should impact credit spreads apart from its influence on policy uncertainty. 14 In recent years, the control of the Senate and the House has switched frequently between the two parties, and presidential election races have tightened, leading to higher uncertainty about government policies (Canes-Wrone and Park (2012)). 15 The updated 2015 database contains DW-NOMINATE scores from the 1st to the 113th Congress. 23