TAXATION AND CORPORATE DEBT: ARE BANKS ANY DIFFERENT?

Transcription

1 National Tax Journal, March 2017, 70 (1), TAXATION AND CORPORATE DEBT: ARE BANKS ANY DIFFERENT? Jost H. Heckemeyer and Ruud A. de Mooij Variation in the responsiveness of firms to corporate tax incentives toward debt finance is important for understanding the presumed effects of the debt bias on macro-financial stability. This holds especially for the difference in responsiveness between banks and non-banks. Using a large cross-country micro panel of consolidated firm accounts, we find relatively large responses for the biggest non-financial companies, although these effects are less pronounced as conditional leverage ratios increase. The smallest effects are found for large banks. Results are largely robust for attenuation bias. Keywords: corporate tax, debt bias, leverage, banks, financial stability, quantile regression JEL Codes: G21, G32, H25 I. INTRODUCTION Tax systems affect corporate financial structures in at least two ways. First, as interest payments are deductible for the corporate income tax while equity returns are not, tax systems encourage the use of debt rather than equity finance debt bias. Second, international corporate tax rate differentials provide an incentive for multinational companies to allocate debt to high-tax subsidiaries through intra-company lending, and have the interest income taxed in low-tax affiliates debt shifting. There is ample empirical evidence of both debt bias and debt shifting. Yet, most empirical studies make no distinction between the two. For instance, two meta studies have examined a wealth of research papers that mostly assess the combined effect of corporate tax rates on (internal plus external) debt ratios. They conclude that, on average, the literature finds that a one percentage point higher corporate tax rate will increase the debt-asset ratio by between 0.2 and 0.3 percentage points (Feld, Heckemeyer, and Overesch, 2013; de Mooij, 2011). Only few studies distinguish between internal and external debt and assess the differential impact between debt bias and debt shifting Jost H. Heckemeyer: Leibniz Universität Hannover, Hannover, Germany (heckemeyer@steuern. uni-hannover.de) Ruud A. de Mooij: International Monetary Fund (IMF), Washington, DC, USA (rdemooij@imf.org)

2 54 National Tax Journal (Altshuler and Grubert, 2003; Desai, Foley, and Hines Jr, 2004; Buettner et al., 2009). Several studies, most notably for the United States, focus only on debt bias. Some others look only at debt shifting by assessing the tax effect on intracompany debt (Overesch and Wamser, 2010; Buettner and Wamser, 2013; Buettner, Overesch, and Wamser, 2016). The conclusion arising from these studies is that debt bias and debt shifting are both important, but possibly to varying degrees, with some indication that internal debt is more sensitive to tax factors than external debt (Feld, Heckemeyer, and Overesch, 2013). Also the welfare implications of debt bias and debt shifting are different. Debt shifting is largely a form of international tax arbitrage by multinational companies. This reduces government revenue and, by requiring higher corporate tax rates, causes indirect welfare losses. However, the direct welfare loss from debt shifting is probably small. Debt bias, in contrast, directly creates allocative distortions e.g., reflected in higher agency costs or bankruptcy costs which result in direct as well as indirect welfare losses (Gordon, 2010; Sorensen, 2014). Views on the social cost of debt bias have changed since the financial crisis. Indeed, many have expressed concerns regarding high corporate debt ratios in light of their implications for financial stability. 1 High levels of external corporate debt, for instance, are thought to increase firm default risks, magnify rollover risk, and exacerbate debt overhang. In this way, high debt ratios raise the probability of financial crisis and magnify its depth. The welfare costs of debt bias might therefore be larger, possibly much larger than often thought, especially when it occurs in the financial sector. In light of these potentially large social costs of debt bias related to macro-financial instability, this paper focuses on debt biases related to external debt. In particular, we use a large cross-country micro panel data set of consolidated financial statements of companies. The data comprise 63 countries, covering the years between 1996 and Since the consolidation of accounts eliminates internal debt, we only assess debt bias and not debt shifting. One possible problem with consolidated data is, however, measurement error in the tax variable. In particular, foreign subsidiaries consolidated in the accounts may not be subject to the tax rate faced by the parent company, e.g., if the parent country employs a territorial system or because tax deferral mitigates homecountry taxation. The domestic tax of the parent then does not correctly measure the incentive for debt bias, which leads to attenuation bias. To explore the robustness of our findings, we compare our main results with similar regressions based on unconsolidated accounting data. The social costs of debt bias may vary among firms, as some firms might be (much) more important for macro-financial stability than others. The main contribution of this paper is to shed light on three dimensions of firm heterogeneity that can be particularly important for the financial stability implications of debt bias, namely: (1) the difference between banks and non-banks; (2) firm size; and (3) initial firm leverage. First, we explore the difference between debt bias in the banking sector and for other firms. Debt bias is potentially a far greater concern in the banking sector than elsewhere 1 This view is not entirely new; see, e.g., Bernanke and Campbell (1988).

3 Taxation and Corporate Debt: Are Banks Any Different? 55 in the economy, as a bank failure can increase the risk of contagion through the financial system and jeopardize the stability of the financial system as a whole. The capital structure of banks might theoretically be governed by different factors than those of non-financial firms. For example, taxes might matter less for banks than for non-banks because capital requirements restrict banks from having too high debt. Yet, banks may face similar tax incentives for debt finance as long as they hold equity in excess of regulatory requirements. Moreover, the corporate governance structure of a bank may differ from other firms because banks enjoy explicit or implicit guarantees from the state associated with deposit insurance or too-big-to-fail status. 2 The difference in the response to tax incentives for debt finance by banks and non-banks is therefore theoretically ambiguous and ultimately an empirical issue. Existing empirical studies on debt bias, however, have largely ignored the financial sector, either by eliminating data on financial firms, or by including them as part of a broader set of unspecified companies. Only recently have researchers started to explore empirically the effects of taxation on debt bias in the financial sector (de Mooij and Keen, 2016; Hemmelgarn and Teichmann, 2014; Gu, de Mooij, and Poghosyan, 2015; Merz and Overesch, 2016). This paper contributes to this literature by directly comparing the tax responsiveness of banks and non-banks, using a single dataset: Worldscope. Second, we look at differences in firm sizes. The bankruptcy of a large firm might induce relatively large spillover effects to other companies and thus exert relatively large effects on macro-financial stability. Theoretically, the tax responsiveness of large firms might be different from small firms. For example, small firms are usually more credit constrained. Yet, small firms also more often seek credit from insiders in the firm, who can be particularly flexible. Large firms, on the other hand, can issue debt securities, i.e., commercial papers or corporate bonds, on the capital market, making them more responsive to tax rates. To explore the relevance of size, we follow Gordon and Lee (2001) by interacting the tax term in our regressions with polynomials of asset size. While Gordon and Lee use aggregated tax-return data for discrete size groups of U.S. companies, our study uses micro data for firms from 63 countries. Finally, we look at the impact of initial leverage on the responsiveness of firms to tax factors. Firms with high initial debt levels might face larger default and rollover risks and therefore can be more important to spreading instability. Tax responses may be smaller at the margin for those companies, since they are more likely to be credit constrained. For a bank, for instance, a high leverage ratio means that it has a smaller buffer relative to the minimum capital requirement and, therefore, possibly less flexibility in adjusting debt. Yet, high leverage may also signal easy access to credit and, therefore, may be associated with a high responsiveness to tax rates. This paper will use quantile regression to explore the relationship between capital structure and taxation at different points of the conditional leverage distribution. 2 See de Mooij and Keen (2016) for a theoretical model of the differences between banks and non-banks that may give rise to differential responsiveness. An important issue for banks is the interplay between taxation and regulation, which is a relatively new area of research; see, e.g., de Mooij and Nicodème (2014).

4 56 National Tax Journal Our results inform the debate on the social costs of debt bias through its impact on financial stability. Specifically, the empirical findings suggest that, on average, banks are less responsive to tax incentives for debt finance than non-banks. For non-banks, we find a U-shaped pattern between asset size and tax responsiveness, implying that the largest responses are observed for very small and very large companies. For banks, however, tax responsiveness declines linearly in asset size over the relevant part of the size distribution. The quantile regressions reveal that large banks are the least responsive to tax, irrespective of their conditional leverage although these estimates might be subject to attenuation bias. Large non-banks, instead, are highly responsive, especially at low conditional leverage ratios. The rest of this paper is organized as follows. Section II discusses our methodology and data. Section III presents the basic empirical results, including the quantile regressions. Section IV assesses the robustness of our findings for clustering assumptions and possible measurement error. Section V concludes. II. METHODOLOGY A. Specification and Estimation This paper estimates the following reduced-form baseline equation that is commonly used in the empirical literature on corporate finance, including for banks (Fan, Titman, and Twite, 2012; Gropp and Heider, 2010; Frank and Goyal, 2009) (1) lev = β + β τ + β x + β z + δ + θ + µ + ε, isct 1 2 ct 1 3 ict 1 4 ct 1 s c t isct where the subscript i denotes each firm, the subscript s denotes each industry, the subscript c denotes each country, and the subscript t denotes each year. The dependent variable, lev, is the book leverage-to-asset ratio, and t is the country s statutory corporate income tax (CIT) rate. 3 The vector x ict 1 includes firm-level variables that typically have an influence on leverage ratios. Consistent with the existing literature (Fan, Titman, and Twite, 2012; Gropp and Heider, 2010; Titman and Wessels, 1988), we include asset tangibility (fixed assets divided by total assets), profitability (operating income divided by total assets), firm size (the natural logarithm of total assets), and the market-to-book ratio (the market value of equity divided by the book value of equity). For non-banks, capital structure regressions typically find that asset tangibility, as a proxy for collateral, increases access to external funding and thus leads to higher leverage ratios. The nature of the banking sector (such as maturity transformation) and the impact of regulation may produce a 3 The seminal work by Miller (1977) suggests that capital structures adjust to the after-personal-tax benefit of debt. The quantitative meta-analysis of Feld et al. (2013) shows that including personal taxes in the tax term generally does not lead to significantly different effect estimates. The capital structure decision, in practice, thus seems to be determined by corporate-level taxation rather than the (unknown) personal tax position of the marginal investor.

5 Taxation and Corporate Debt: Are Banks Any Different? 57 different relationship for banks, however, i.e., most assets will be in the form of loans, not tangible assets. As to firm size, most studies find that larger firms feature higher leverage ratios. Higher profits add to equity when retained within the firm and thus directly reduce the leverage ratio. Yet, profitability may also signal good health for a firm implying easier access to credit, which could raise the debt ratio. A priori, the effect of profitability is thus ambiguous. With respect to the market-to-book ratio, Fan, Titman, and Twite (2012) argue that this variable captures the effects of other factors, such as operating risk, growth opportunities, R&D expenditure, capital expenditure, selling expenses, and collateral value of assets, which are difficult to directly control for due to data limitations. In their in-depth assessment of the determinants of bank capital structure, Gropp and Heider (2010) indeed find that the market-to-book ratio co-moves strongly with other measures of asset risk. 4 Equation (1) includes a full set of industry fixed effects d s (for the regression of non-banks) and country fixed effects q c. Industry effects were assessed by MacKay and Phillips (2005) and were confirmed to be important by Feld, Heckemeyer, and Overesch (2013) and Frank and Goyal (2009). Country fixed effects account for time-invariant country-specific features that may determine capital structures. These include, for example, legal factors, i.e., the bankruptcy code, and matters of capital supply (Demirgüç-Kunt and Maksimovic, 1999; de Jong, Kabir, and Nguyen, 2008; Fan, Titman, and Twite, 2012). In addition, the vector z ct 1 includes the growth rate of GDP and the rate of inflation to allow for country-specific variation in macro variables over time. High growth at the country level has no clear theoretical implications for debt ratios and might also control for factors correlated with high growth, such as credit growth. Inflation may lead to higher risk premiums, as it may reveal uncertainty about future price developments and thus unexpected inflation, and discourage debt supply. Yet, as nominal interest is deductible for the CIT, high inflation may also encourage debt finance as it lowers real borrowing costs. Lastly, we add a full set of year dummies to absorb common shocks over time affecting capital structures across countries and firms. Consistent with Fan, Titman, and Twite (2012), all firm-level and country-level variables are lagged one period to allow for the non-contemporaneous nature of the relationship between such characteristics and financing choices. We estimate (1) as our baseline specification. We also augment (1) with interactions of (polynomials of) firm size and the CIT rate to explore tax effects along the size distribution of firms. Moreover, robustness checks are performed for the inclusion of control variables, such as proxies for the regulatory environment of banks, including deposit insurance and minimum capital requirements (see discussion below). 5 4 Specifically, Gropp and Heider (2010) report a correlation coefficient of 0.85 between the market-to-book ratio and asset risk, where the latter is defined as annualized standard deviation of (daily stock price returns) (market value of equity/market value of bank). 5 We do not include thin capitalization rules in the regressions for two reasons. First, these rules generally do not apply to banks. Second, two-thirds of all countries employing such rules, apply them only to intracompany debt (while our interest is in external debt).

6 58 National Tax Journal To explore the differences between banks and non-banks, we estimate a full dummyinteractive model, i.e., we include interaction terms of all covariates, including the constant, with a dummy that distinguishes banks from non-banks. This is statistically equivalent to estimating two different samples, but it allows us to directly test whether the tax effects differ between banks and non-banks. While ordinary least squares (OLS) regressions estimate the conditional mean response of financial leverage to taxes, we also present quantile regressions, which show non-central locations of the response distribution and thus provide a more complete understanding of how these are affected by the explanatory variables (Hao and Naiman, 2007). 6 We will estimate models for the conditional median and the quantiles 0.05, 0.10, 0.25, 0.75, 0.9, and Consistent with Koenker and Hallock (2001), we report and plot graphs of the tax coefficient estimates from the quantile regressions to show how tax effects vary across the conditional leverage distribution. B. Data Firm-level data are taken from Worldscope, compiled by Thomson Financial. The database contains financial information on public companies from a wide range of industries, countries, and years. We use data from 63 countries for the 15 years between 1996 and After excluding observations with missing values for required variables, and given that we estimate specifications with lags, the full estimation sample consists of 31,846 firms and a total of 229,608 firm-year observations; 30,432 non-banks contribute 217,144 of these observations. We identify banks using the Industry Classification Benchmark (ICB) jointly created by FTSE and Dow Jones. ICB groups companies into 10 industries and derives from these in increasingly finer classifications 19 supersectors, 41 sectors, and 114 subsectors. Each public company in the investable stock universe is uniquely classified, based on the company s primary revenue source. We refer to the sectoral classification, which, within the financial industry, distinguishes banks from seven other sectors such as insurance, real estate investment, and other financial services (e.g., asset management). According to the ICB definition, banks provide a broad range of financial services, including retail banking, loans, and money transmissions. 7 We have 1,414 banks in the sample with 12,464 bankyear observations. The banks share in the total number of observations is 5.7 percent. The Worldscope data reflect the consolidated accounts of firms so that debt ratios measure the external debt of the corporation. 8 Many earlier studies have instead used 6 Quantile regression was introduced by Koenker and Bassett (1978) and can be viewed as an extension of classical least squares estimation. Instead of approximating the conditional expectation function E[Y x], quantile regression estimates conditional quantile functions. The central special case is the median regression, which minimizes a sum of absolute errors, instead of squared errors as with OLS. Other conditional quantile functions are estimated by minimizing an asymmetrically weighted sum of absolute errors (Koenker and Hallock, 2001). 7 For more information on the chosen industry classification, please see ICBDocs/Structure_Defs_English.pdf. 8 The data only include public companies and results cannot be generalized toward privately-held companies.

7 Taxation and Corporate Debt: Are Banks Any Different? 59 unconsolidated accounting data, which cannot distinguish between external and internal debt and therefore between debt bias and debt shifting. Given our interest in debt bias as a concern for macro-financial stability, we use consolidated statements. This, however, may cause potential measurement error in the tax variable, as foreign branches or wholly-owned foreign subsidiaries face different tax rates from the parent. One way to address this attenuation bias is by using an alternative tax variable that takes account of the different tax rates applying to foreign subsidiaries. For instance, one could use the actual average tax paid to measure the tax burden. Yet, the average tax rate would be endogenous as it depends on the leverage ratio itself. An alternative is to use unconsolidated accounting data. These may suffer less from measurement error in the tax variable, although they might not be free from it either, e.g., as banks can own (foreign) branches that are taxed on their worldwide income (and branches as an organizational structure instead of subsidiaries are particularly prevalent in the banking sector) or to the extent that firms reside in countries with worldwide tax systems. 9 Yet, unconsolidated data do not usually distinguish between internal and external debt so that the regressions have a different interpretation, i.e., they measure the combined effect of debt shifting and debt bias. It is nevertheless interesting to see if our main results based on consolidated accounts carry over to regressions based on unconsolidated accounts, which we run in a separate sensitivity analysis. The Worldscope data reveal important differences between banks and non-banks. Figure 1 shows the histograms and cumulative distributions of size, in terms of the logarithm of total assets, of non-banks (Figure 1a) and banks (Figure 1b). The log size distribution of the log of total assets of non-banks is close to normal. Some 25 percent have a log of total assets below ($45.9 million) whereas the 75 percent quantile is at ($648.8 million). 10 The 1 percent and 5 percent quantiles are, respectively, at and ($1.4 million and $6 million). The 95 percent and 99 percent quantiles are, respectively, at and ($7.2 billion and $39.1 billion). Banks in our sample are generally larger than non-banks. Specifically, we see a skewed size distribution for banks, with only few small banks. The 1 percent and 5 percent quantiles are at, respectively, and ($73 million and $148 million). Some 25 percent have a log of total assets below ($509.7 million) whereas the 75 percent quantile is at 23.3 ($13.1 billion). The right tail is rather long with numerous large and very large banks in the sample. The 90 percent, 95 percent, and 99 percent quantiles are, respectively, 24.7, 25.75, and ($53.2 billion, $153.1 billion, and $989.1 billion). Table 1 presents summary statistics of the main variables used in the analysis. Company-level variables, including size, are winsorized at the 1 percent and 99 percent levels. The table shows descriptive statistics for, respectively, the full sample, the subsample of banks, and the subsample of non-banks. We observe large differences between banks and non-banks with respect to the leverage ratio (defined as the ratio of total liabilities to total assets of the firm): the mean leverage ratio of banks exceeds 90 9 Empirical evidence shows that U.S. multinationals often avoid repatriation taxes by income shifting or by deferring repatriation of income (Foley et al., 2007). 10 All monetary values in the figures and tables are in U.S. dollars in this paper.

8 60 National Tax Journal Figure 1 Size Distribution of Non-Banks (A) and Banks (B) Fraction K 10K 100K 1M 10M 100M 1B 10B 100B 1T A. Non-Bank Total Assets (on Log Scale) ($) Cumulative Sample Distribution Fraction K 10K 100K 1M 10M 100M 1B 10B 100B 1T B. Bank Total Assets (on Log Scale) ($) Cumulative Sample Distribution percent, while that of non-banks is less than 50 percent. Tangibility (the share of tangible assets in terms of total assets) is less than 2 percent on average for banks, whereas it is over 30 percent for non-banks. Appendix A provides more information about the sample by country. 11 III. RESULTS A. Ordinary Least Squares Table 2 shows the results from our main capital structure regression, estimated by OLS. The first three columns of Table 2 show the regression results for non-banks, whereas the last four columns refer to banks. The first specification is our baseline specification in (1). The second adds a simple linear interaction of firm size with the CIT rate. The third allows for more flexibility by interacting the tax variable with a polynomial of size. We add powers to the polynomial as long as each additional power 11 The online Appendix is available at Heckemeyer/HDM_OA.pdf.

9 Taxation and Corporate Debt: Are Banks Any Different? 61 Table 1 Descriptive Statistics Sample Mean Standard Deviation Median Minimum Maximum Leverage ratio All firms Banks Non-banks Corporate income tax rate All firms Banks Non-banks Log total assets All firms Banks Non-banks Profitability All firms Banks Non-banks Market-to-book All firms Banks Non-banks Tangibility All firms Banks Non-banks GDP growth rate All firms Banks Non-banks Inflation rate All firms Banks Non-banks Note: The number of observations is 229,608 for the full sample including all firms, 12,464 for the sample with banks, and 217,144 for the sample with non-banks. Firm-level variables are winsorized at 1 and 99 percent levels. is statistically significant. For both non-banks and banks, we arrive at a quadratic functional relationship. For banks, we also show a fourth specification in which we add two regulatory variables to the regression: deposit insurance and capital requirements. In all regressions, our statistical inference controls for clustering of observations at the firm level. Before discussing the effects of taxation, we first consider the impact of non-tax factors on capital structures. First, for non-banks Columns 1 3 of Table 2 show that firm size exerts a statistically significant positive effect on leverage. Profitability enters with a statistically significant negative coefficient, reflecting the rise in equity as a result of higher retained earnings. Tangibility has a positive coefficient as expected,

10 62 National Tax Journal Table 2 Capital Structure Regressions for Non-Banks and Banks (1) (2) (3) (4) (5) (6) (7) Non-Banks Banks CIT rate 0.169*** 0.439*** 3.049*** 0.059*** 0.855*** 3.326*** 3.693** (6.75) ( 3.21) (3.69) (3.14) (4.84) (2.66) (2.38) Ln assets 0.036*** 0.025*** 0.190*** 0.005*** 0.018*** 0.128*** 0.143*** (58.55) (10.55) (6.41) (8.58) (6.15) (3.26) (2.83) (Ln assets) *** 0.002*** 0.003** ( 5.60) ( 2.86) ( 2.53) Ln assets CIT 0.031*** 0.339*** 0.034*** 0.239** 0.277** (4.53) ( 3.95) ( 4.50) ( 2.29) ( 2.11) (Ln assets) 2 CIT 0.010*** 0.004* 0.005* (4.37) (1.93) (1.85) Profitability 0.144*** 0.144*** 0.154*** * ( 23.43) ( 23.47) ( 24.17) ( 1.49) ( 1.45) ( 1.68) ( 1.38) Tangibility 0.063*** 0.063*** 0.062*** (12.33) (12.41) (12.18) ( 0.86) ( 0.83) ( 0.67) ( 0.13) Market-to-book 0.007*** 0.007*** 0.007*** (23.16) (23.24) (23.81) (1.16) (1.13) (0.85) (1.34) GDP growth 0.068*** 0.071*** 0.071*** (3.84) (4.02) (4.01) (0.25) ( 0.21) (0.01) ( 0.91) Inflation * 0.057* 0.052* 0.073** ( 1.20) ( 1.09) ( 0.97) ( 1.93) ( 1.93) ( 1.81) ( 2.03) Capital requirement Deposit insurance 0.001** ( 2.10) 0.001** ( 1.98) Constant 0.316*** 0.109** 1.666*** 0.761*** 0.474*** 0.828* ( 10.67) ( 2.03) ( 5.83) (38.97) (6.96) ( 1.75) ( 1.61) Number of observations 217, , ,144 12,464 12,464 12,464 10,938 R Note: The table shows regressions of the book leverage-to-asset ratio on firm- and country-level determinants of leverage. All regressions include full sets of time dummies and country fixed effects. Regressions 1 3 for non-banks include full sets of industry dummies. All regressions are from OLS estimation. T-values based on standard errors robust to clustering at the firm-level are in parentheses. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels.

11 Taxation and Corporate Debt: Are Banks Any Different? 63 significant at the 1 percent level. Also the market-to-book ratio has a positive and highly statistically significant effect on the leverage ratio. GDP growth is associated with an increase in corporate debt, while the effect of inflation is negative but statistically insignificant. In regressions 4 7 in Table 2 for banks, some non-tax factors enter with a different sign as compared to non-banks. Size itself enters with a positive coefficient, as is the case for non-banks. Profitability, however, does not exert a robust statistically significant effect on bank leverage. Tangibility enters with a statistically insignificant negative coefficient. Yet, the tangibility variable has a different interpretation for banks than for non-banks. Market-to-book and GDP growth enter with insignificant coefficients for banks. The negative effect of inflation on bank leverage is significant at the 5 percent or 10 percent confidence levels, whereas the inflation effect on non-bank capital structure is less pronounced and insignificant. Next, we turn to the tax effects. In the baseline regressions (Column 1 for non-banks and Column 4 for banks), the tax effect on the leverage ratio is positive and statistically significant at the 1 percent confidence level. The marginal tax effect of for banks is smaller than the coefficient of for non-banks. Columns 2 and 5 show, for non-banks and banks respectively, how tax effects vary along the size distribution. Interestingly, the interaction between tax and size in Column 2 is positive and significant, which indicates that the tax sensitivity of non-bank leverage is increasing with firm size. In contrast, the interaction term in Column 5 is negative and statistically significant, which suggests a decreasing relationship between the marginal tax effect and bank size. Size thus exerts opposite effects on the tax responsiveness of banks and non-banks. Regressions with quadratic size interactions are shown in Columns 3 and 6 of Table 2. All tax terms are statistically significant at the 1 percent level. The positive coefficient for the quadratic interaction of size and the CIT rate implies a convex relationship between the marginal tax effect and size, for both non-banks and banks. Using these regression results, Figures 2 and 3 show how the marginal tax coefficient (on the vertical axis) varies along the size distribution (on the horizontal axis), for nonbanks and banks, respectively. In both figures, the black line represents point estimates, while the grey lines show the 95 percent confidence intervals. The figures also depict histograms that illustrate the fraction of firms that fall in the different size bins. Figure 2 shows that non-banks feature a moderate tax sensitivity over the most relevant range of the size distribution, with marginal tax effects ranging between 0.1 and 0.2. The relatively large response for small firms (a tax effect of 0.2 at total assets of $1.3 million) may reflect a high tax elasticity of insider funding, which is generally used extensively by small firms. From the left tail almost to the center of the distribution, the tax sensitivity decreases with firm size. Non-banks with total assets of around $40 million feature the smallest tax sensitivity, with a marginal tax effect on leverage close to 0.1. Toward the right tail of the firm size distribution, tax responsiveness rises steeply. Large firms thus seem to have more flexibility in adjusting debt, possibly reflecting easier

12 64 National Tax Journal Figure 2 Marginal Tax Effect on the Leverage Ratio along the Size Distribution of Non-Banks M 10M 100M 1B 10B 100B 1T Total Assets (on Log Scale) ($) Figure 3 Marginal Tax Effect on the Leverage Ratio along the Size Distribution of Banks M 10M 100M 1B 10B 100B 1T Total Assets (on Log Scale) ($)

13 Taxation and Corporate Debt: Are Banks Any Different? 65 access to credit. If total assets exceed $10.5 billion, the marginal tax effect is larger than 0.4. These results are similar to those in Gordon and Lee (2001). Using aggregate firm data from tax returns with firms classified by 14 different asset-size intervals, they find a U-shaped pattern between taxes and leverage along the size distribution. We find the same relationship based on micro data. Contrary to non-banks, the tax sensitivity of banks is decreasing continuously along the relevant size distribution (Figure 3). Relatively small banks with total assets of $100 million show a pronounced marginal tax response, reaching a value of Marginal tax effects for banks with total assets above $26 billion are small and no longer significantly different from zero. The increasing portion of the quadratic relationship appears to be irrelevant, as it applies to an area where no banks exist. This result is consistent with de Mooij and Keen (2016), who find that for the largest 5 percent of banks in their sample, tax effects are considerably smaller than for the average bank. Column 7 of Table 2 augments the specification of Column 6 by adding two regulatory variables: deposit insurance and capital requirements. Deposit insurance reduces the risk of bank runs and thus the banks need to hedge and seek more liquid sources of finance (Fan, Titman, and Twite, 2012). In particular, banks might increase their leverage in order to maximize the subsidy from incorrectly priced deposit insurance (Gropp and Heider, 2010). Column 7 of Table 2 uses data on deposit insurance coverage limits as a percentage of deposits per capita, compiled by Demirgüç-Kunt, Karacaovali, and Maksimovic (2005), which are available for years up to For years after 2003, we assume constant deposit insurance levels. Capital requirements may reduce the flexibility of banks to adjust their capital structure. Yet, theories of optimal bank capital structure suggest that capital requirements are not necessarily binding (Flannery, 1994; Myers and Rajan, 1998; Diamond and Rajan, 2000; Allen, Carletti, and Marquez, 2011). To capture the degree of regulatory intervention, we employ the Capital Regulatory Index extracted from the World Bank surveys on bank regulation covering the years 1999, 2003, 2007, and 2011 (see Barth, Caprio, and Levine, 2013). For sample years not covered by the index, we use the information from the next available subsequent year. The index incorporates several dimensions of capital stringency, such as the regulatory capital adequacy regime used, the risks covered by the minimum capital requirement, and the extent to which market value losses are deducted from capital before minimum capital adequacy is determined. 12 As both regulatory variables are not available for the full set of countries in our sample, the number of observations used in the regression drops to 10,938. The estimation results in Column 7 of Table 2 show that capital regulation stringency exerts a negative influence on banks debt levels, as one would expect. The effect is significant at 5 percent confidence. Deposit insurance enters with an unexpected negative coefficient, which is also significant at 5 percent. As in Gropp and Heider (2010), we thus do not find evidence for deposit insurance reducing banks capitalization. Importantly, however, we find that the inclusion of regu- 12 Basel II requires that the sum of Tier 1 and Tier 2 capital be at least 8 percent of risk-weighted assets. Basel III requirements developed since 2008 are more demanding, but lie outside our sample period. The minimum capital ratio in Barth et al. (2013) varies between 6 and 15 percent, with an average of 8 percent. Table A2 in Appendix A shows Capital Regulatory Index values across countries and over time.

14 66 National Tax Journal latory variables hardly affects the coefficients of the tax terms in the capital structure regressions. 13 To test formally whether tax effects for non-banks and banks differ along the size distribution, we use the specification in Columns 3 and 6 of Table 2 to compare marginal tax effects at seven quantiles of the bank size distribution. In particular, Table 3 juxtaposes the tax effects found for banks for quantiles 0.05, 0.10, 0.25, 0.5, 0.75, 0.9, and 0.95 of the bank size distribution and equally sized non-banks. An F-test is used to verify whether marginal tax effects differ. The results reported in Table 3 show that we cannot reject the null hypothesis of no difference in tax effects (p-value of 0.11) between non-banks and banks at the first quartile (Q25) of the bank size distribution ($509.7 million of assets). The tax effect for smaller banks (Q5 and Q10) is found to be significantly larger than that for small non-banks; the tax effect for large non-banks is significantly larger than for banks (Q50 and higher). B. Quantile Regressions The tax responsiveness of debt might not only vary over the size distribution, but also over the leverage distribution. For instance, capital-constrained banks might respond less to taxes than capital-abundant banks in light of the smaller buffer they hold relative to the minimum capital requirement. The conditional mean effect estimated by OLS may thus reflect responses that come primarily from capital-abundant banks. Quantile regressions can shed light on whether the initial leverage ratio matters for the responsiveness to tax. The quantile regressions focus only on the most flexible specification, i.e., the specification with an interaction of the tax rate with a second-order polynomial of size. We look at the 0.05, 0.10, 0.25, 0.5, 0.75, 0.9, and 0.95 quantiles of the conditional leverage distribution. Given the importance of firm size, we evaluate each quantile regression at three asset levels, namely, the 0.1, 0.5 and 0.9 quantile of the respective asset distribution. Tables A3 and A4 in Appendix B show the regression outcomes. We focus only on the tax effects, which are shown in Figure 4 and Figure 5. The three lines in each figure show, for each of the three size categories, the tax effects along the conditional leverage distribution (on the horizontal axis). Figure 4 shows the results for non-banks. For large firms, the responsiveness to tax rates declines with higher conditional leverage ratios. For relatively small firms, however, the pattern is the opposite: the responsiveness to tax rates rises as conditional leverage ratios grow larger. For the median firm, tax sensitivity is fairly flat. Accordingly, the biggest tax responses are found for large firms with a low to median conditional leverage ratio and for relatively small firms with a relatively high conditional leverage ratio. The smallest tax responses are found for relatively small firms with 13 Specifically, the curve of joint marginal tax effects along the bank size distribution according to results in Column 7 of Table 2, available from the authors upon from request, is all but identical to the one shown in Figure 2.

15 Taxation and Corporate Debt: Are Banks Any Different? 67 Table 3 Marginal Tax Effects of Non-Banks and Banks along the Bank Size Distribution Banks Q5 Banks Q10 Banks Q25 Banks Q50 Banks Q75 Banks Q90 Banks Q95 Bank Size ($Million) ,900 13,100 53, ,100 Non-Banks 0.108*** 0.122*** 0.156*** 0.242*** 0.424*** 0.603*** 0.763*** Banks 0.324*** 0.289*** 0.231*** 0.144*** 0.047** F-Test (P>F) (0.0015) 7.89 (0.0050) 2.46 (0.1166) 6.87 (0.0087) (0.0000) (0.0000) (0.0000) Notes: The table shows the joint marginal tax effects on the leverage-to-assets ratio from Columns 3 and 6 of Table 2. Joint effects are computed as linear combinations of the estimated CIT coefficient and the coefficients of the interactions between tax and firm size, evaluated at different points of the sample distribution of firm size (log of total assets). To explore the differences between banks and non-banks, we estimate the two specifications for non-banks and banks in a full dummy-interactive model, i.e., we include interaction terms of all covariates, including the constant, with a dummy that distinguishes banks from non-banks. Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels based on standard errors robust to clustering at the firm-level. The F-tests test for the difference between joint tax effects for banks and non-banks. P-values are given in parentheses.

16 68 National Tax Journal Figure 4 Tax Effects from Quantile Regression for Three Size Categories of Non-Banks Quantiles of the Conditional Leverage Distribution Coefficient Evaluated at p10 of Non-bank Size Coefficient Evaluated at p50 of Non-bank Size Coefficient Evaluated at p90 of Non-bank Size Figure 5 Tax Effects from Quantile Regression for Three Size Categories of Banks Quantiles of the Conditional Leverage Distribution Coefficients Evaluated at p10 of Bank Size Coefficients Evaluated at p50 of Bank Size Coefficients Evaluated at p90 of Bank Size

17 Taxation and Corporate Debt: Are Banks Any Different? 69 relatively low conditional leverage ratios and large firms with high conditional leverage ratios. Figure 5 shows that the tax response declines with the conditional leverage ratio for small and median-sized banks. At the 0.95 quantile of the conditional leverage distribution, the tax effect on leverage is, respectively, 0.03 and 0.07, both significant at the 1 percent level. For large banks, the tax response hardly varies along the conditional leverage distribution. Moreover, it is always smaller than the response for small and median banks. Indeed, the tax responses of large banks are always insignificant, irrespective of leverage. An F-test suggests that at the 0.95 quantile of conditional leverage, the response of large banks remains significantly lower than that of small and median-sized banks. These findings are important, as the largest banks tend to be most important for the overall stability of the financial system. Some caution is needed, however, since large banks might have relatively more foreign operations than small banks, so that measurement error in the tax variable is potentially more problematic (see the next section). Also, the positive responses by highly leveraged small and median banks is a cause of concern in light of contagion risks. IV. ROBUSTNESS This section explores the robustness of our findings with respect to two possible econometric pitfalls: (1) the assumption regarding the clustering of standard errors; and (2) bias due to measurement error. A. Clustering of Standard Errors at the Country Level The first issue relates to statistical inference. Specifically, correlation of regressors within countries might produce a downward bias in the estimated standard errors. Our main interest is on tax-size interactions which should not be highly correlated between firms within countries. However, as the CIT rate itself only varies at the country level, we check whether our main findings change when using standard errors that are robust to clustering at the country level. Ultimately, the results turn out to be robust. More specifically, the tax coefficient in the baseline regression 1 of Table 2 for nonbanks remains significant at the 1 percent level, if statistical inference is based on standard errors robust to clustering at the country level. Moreover, the positive coefficient of the quadratic size interaction with the tax rate in the flexible regression (3) remains significant at the 10 percent level. However, the coefficient of the linear size interaction term in that specification is no longer significant. Still, the convex relationship shown in Figure 2 is robust to clustering at the country level. 14 Considering the results for banks, the tax effect on the leverage ratio in the baseline regression, i.e., Column 4 of Table 2, remains significant at the 5 percent level with standard errors robust to clustering at the country level. In the flexible regression (6) of 14 We do not repeat coefficient estimates from Table 2 but changes in t-values are as follows: With standard errors robust to clustering at the country level, the t-value of the CIT coefficient in regression 1 of Table 2 drops to 3.10, and t-values of the coefficients of the linear and quadratic size interaction with tax in regression 3 of Table 2 are, respectively, 1.50 and 1.68.

18 70 National Tax Journal Table 2 for banks, the linear size interaction again remains significantly negative, but the quadratic size term becomes insignificant. It suggests that the linear and downward sloping relation with bank size, as shown in Figure 3, might hold across the entire size distribution. 15 B. Measurement Error Measurement error in the tax variable may occur because consolidated financial accounts include foreign earnings that might face a different tax rate than that of the parent. This can lead to estimates that are biased toward zero. The size of this possible attenuation bias will depend on the propensity of multinational activity, which may differ between banks and non-banks. For instance, to the extent that banks operate relatively more domestically, the bias for banking will be notably smaller. However, larger banks might have more foreign operations and for them the bias might be larger. To examine the robustness of our findings to this possible attenuation bias, we employ two alternative datasets with unconsolidated accounts, performing the same regressions as in the previous section. For non-banks, we use unconsolidated firm-level data from Amadeus, provided by Bureau van Dijk. For banks, we use unconsolidated accounting data from Bankscope, also provided by Bureau van Dijk. Unlike Worldscope, the Amadeus data are only available for European countries and years. We therefore perform the robustness analysis for a sample of 35 European countries between 2001 and The Amadeus data comprise 16,499,901 firm-year observations. To compare the results for non-banks with those for banks, we sample unconsolidated bank data again for all available European countries. However, here we also add the United States. 16 Using the same years , we obtain a dataset of 89,997 bank-year observations. 17 We use the same empirical specification as in (1) and again augment it with interactions of (polynomials of) firm size and the CIT rate. 18 For banks, we construct an alternative, more inclusive indicator for collateral, following Gropp and Heider (2010), which is defined as the value of total securities plus cash and due from banks divided by total assets. Some control variables are excluded, such as the market-to-book ratio 15 Again, we do not repeat coefficient estimates from Table 2 but provide changes in t-values: With standard errors robust to clustering at the country level, the t-value of the CIT coefficient in regression 4 of Table 2 for banks drops to 2.25; t-values of the coefficients of the linear and quadratic size interaction with tax in regression 6 of Table 2 are, respectively, 1.81 and Adding a large number of data from the United States does not contribute significant time variation in tax rates but helps isolating non-tax influences on bank capital. 17 Descriptive statistics and the list of countries included in unconsolidated non-bank and bank samples are given in Tables A5, A6, and A7 in Appendix C. To verify the importance of the geographical restriction in data, we re-estimated the regressions of Table 2 with the consolidated Worldscope data, using the same group of countries and years as used in Table 4. Our main findings turn out robust; results are available upon request. 18 Size might have a different meaning in this case, since it refers to the size of the subsidiary rather than that of the entire group.

19 Taxation and Corporate Debt: Are Banks Any Different? 71 (which is unavailable for unconsolidated accounts). Columns 1 and 2 of Table 4 show the regression results for non-banks; Columns 3, 4, and 5 refer to banks. The first column of Table 4 shows the regression for non-banks without size interactions. The tax coefficient is and highly significant. This effect is in line with previous empirical literature (Feld, Heckemeyer, and Overesch, 2013; de Mooij, 2011) and twice as large as the effect reported in Column 1 of Table 2. This larger effect has two potential explanations. One is that measurement error in the consolidated accounting data indeed leads to downward bias in estimated effects in Table 2. Another is that unconsolidated data capture effects through both external and intra-firm debt, with the latter being relatively more tax sensitive (consistent with Desai, Foley, and Hines Jr. (2004) and Feld, Heckemeyer, and Overesch (2013)). From our data, it is not possible to choose between these two explanations. Column 2 of Table 4 adds the interaction of the tax variable with a polynomial of size. We add powers as long as they are statistically significant and arrive at a third-order polynomial. Figure A1 in Appendix C illustrates the joint marginal tax coefficients along the entire size distribution of non-banks. It shows that, toward the lower end of the size distribution, around total assets of $100,000, marginal tax effects are large. 19 Given that we use unconsolidated data from both public and privately held firms, many firms indeed feature total assets below $1.4 million, the minimum of the consolidated non-bank sample. Toward the higher end of the size distribution, the tax sensitivity decreases and firms with total assets ranging between $50 million and $100 million feature the lowest tax sensitivity. Above this range, the tax response steeply increases, reaching up to 0.4 for the largest firms. While the third-order polynomial using the unconsolidated data thus gives a somewhat different pattern between tax-response and firm size as compared to what we find for the consolidated data, it is remarkably similar for the comparable size range between $1 million and $1 billion. The main difference is that the response for the small (i.e., at around $100,000) and the largest firms (around $1 billion) is somewhat bigger. To compare the tax responsiveness of non-banks to the one of banks, we consider tax effects also at central quantiles of the bank size distribution (see the bottom of Column 2 in Table 4). A non-bank featuring the size of a Q25 bank displays a tax sensitivity of 0.076, a non-bank as large as a median bank features a marginal tax effect of 0.12 and of at the Q75 of the bank size distribution. Column 3 of Table 4 reports a small and insignificant tax coefficient of for banks. This is smaller and less significant than the coefficient in Table 2. Column 4, however, shows that the difference in results between consolidated and unconsolidated data is effectively small. In this case, a linear interaction with bank size is used as higher orders of size are insignificant. The negative coefficient for the tax-size interaction is similar to what we observe in Figure 3 for consolidated data. Looking at the joint tax effect at the 19 For the very small firms in the sample, tax responsiveness declines to almost zero (see Figure A1 in Appendix C).