Does corporate taxation affect cross-country firm leverage?

Transcription

1 Does corporate taxation affect cross-country firm leverage? Antonio De Socio and Valentina Nigro * December 2011 Abstract The aim of this paper is to evaluate the relation between firm leverage and taxation of corporate income using a dataset of European firms, most of them unlisted. Our results suggest that taxation has a positive impact on corporate leverage and explains part of the cross-country variability. Results remain significant in different subsamples. Consistently with the theory of debt tax shield, the positive correlation between debt and taxation is stronger for highly profitable firms. These findings are robust to different estimation methods and to the use of different proxies for the degree of financial development and the characteristics of the legal system of the country where firms are located. JEL Classification: G32, H32. Keywords: leverage, corporate taxation, financial structure. Bank of Italy, Structural Economic Analysis Department; contacts: antonio.desocio@bancaditalia.it; valentina.nigro@bancaditalia.it The view expressed in this article are those of the authors and do not necessarily reflect those of the Bank of Italy. 1

2 1. Introduction Firms require financial resources to run ordinary activities and to invest. Their financial structure is influenced not only by firm characteristics, but also by a number of institutional and macroeconomic factors such as taxation, bankruptcy laws and creditor protection, size and structure of the financial system. The aim of this paper is to evaluate the cross-country relation between firm leverage and taxation of corporate income using a dataset of European firms, most of them unlisted, from Amadeus database. In our study we control for the effect of legal and financial system and for differences in firm characteristics. The literature on capital structure developed after the seminal work of Modigliani and Miller (1958) focused on the simplifying assumptions underlying their irrelevance proposition. The main market imperfections introduced were the role of taxation, the cost of bankruptcy, the agency problems, and the asymmetric information. They are generally related to two stylized views on corporate structure: trade-off theory and pecking order theory. According to the trade-off theory, firms choose the optimal leverage after a comparison of the losses and the gains obtained with debt or equity. On the one side, corporate tax may offer a debt tax shield owing to the deductibility of interest and an incentive to increase debt. On the other side, there are direct and indirect bankruptcy costs of debt, which are mainly related to agency costs between shareholders and debt holders and between shareholders and managers. 1 The pecking order theory is connected to problems of asymmetric information and its main result is the non-existence of an optimal value of leverage. According to this theory, firms prefer internal financing to external sources; among external financing, debt is preferred to equity. 2 The existence and the magnitude of a tax effect on firm leverage have been investigated by an extensive body of applied corporate finance literature. The results of these studies are not conclusive and vary at large according to two key empirical issues. The first one is the nature of the indicator measuring the impact of taxation. 3 The second issue concerns the characteristics of the firms included in the sample. For instance, using data of listed or unlisted firms can lead to very different findings because listed firms can raise capital in a simpler way owing to lower agency problems and asymmetric information. Among other papers that study cross-country differences in leverage of listed firms, Rajan and Zingales (1995) consider companies of the G-7 countries and find that whether taxation have or does not have explanatory power on leverage is highly sensitive to assumptions about the marginal investor tax rate (e.g. if the investor is tax-exempt or is taxed at the top rate). de Jong et al. (2008) use the effective tax rate (defined as taxes over pre-tax income) and find no relation between taxation and debt measures in a sample of companies from 42 countries. Fan et al. (2011) study the impact of institutional factors on leverage of firms from 39 countries and show that taxation, 1 The drawbacks of debt include: a) the risk shifting (Jensen and Meckling, 1976), which is the incentive for levered firm to overinvest in risky projects; b) the underinvestment problem (Myers, 1977), which happens when a highly levered firm passes up an investment with positive net present value, so that high-growth (and high-investing) firms should have higher equity to avoid this situation; c) the behaviour of shareholders, who could extract value from firm at the expense of debt holders through higher dividends. Also there is the free cash flow problem (Jensen, 1986), which is the possibility that managers of highly profitable firms invest in empire building or in unprofitable projects. In this way, debt represents a limit to the amount of cash flow at their disposal. These agency problems may be mitigated by ownership structure: firms that are strictly controlled by few shareholders have less agency costs, because owners are more interested in avoiding bankruptcy. 2 The order depends on the relative cost of the different sources (Myers and Majluf, 1984; Myers, 1984). A related theory evaluates the relationships between old and new investors and the importance of signalling effects when new securities are issued (Ross, 1977). Firms usually sell new equity when the stock value is overpriced (so that the value of shares drops after emissions), while the issue of debt signals that the firm is profitable and can borrow money. 3 Many different indicators have been used like effective tax rate, marginal tax rate, corporate statutory tax rate, or Miller index. We analyse them in details in section

3 measured using Miller index - which includes, in addition to the corporate tax rate, also the personal tax on interests and dividends - has a positive effect on leverage in developed countries, but not in emerging economies. Other studies investigate the relation between debt level and firm-country specific characteristics for unlisted corporations using European data from Amadeus database. In general, they find that the traditional corporate finance theory developed to be applied to large listed firms holds also for smaller companies. Giannetti (2003) finds that taxation, measured with non-debt tax shield (i.e. depreciation of assets, investments tax credit, and R&D expenses), does not have any impact on leverage. Two other papers use a sample of manufacturing firms to specifically evaluate the role of taxation on leverage. Bartholdy and Mateus (2008) show that corporate statutory tax rate is positively related to leverage. Pfaffermayr et al. (2008) focus on the impact of age in the relation between leverage and corporate statutory tax rate, showing that debt tax shield is more important for older companies. Even if these two last papers find evidence of a relation between corporate taxation and leverage, they do not consider the role of the financial markets (both) nor the effect of the legal system (the latter). In this paper we focus on a measure of leverage which includes only financial debts, thus excluding other liabilities like trade debts, following the reasons presented in Welch (2011). We rely on corporate statutory tax rate to evaluate the impact of debt tax shield on firm financial structure. Therefore we do not have the endogeneity problem related to the effective tax rate nor the need to estimate future income of firms to obtain the marginal tax rate. Moreover we consider benefits and disadvantages of debt just from the point of view of a firm, to avoid the limitations and the measurement problems of Miller index. We include other variables to control for country-level characteristics such as legal system and development of financial markets. Firm individual features are also taken into account; in particular, since we are interested in financial debt only, we check for the effect of net working capital, which in some countries requires a significant amount of funding. Our results suggest that taxation has a positive impact on corporate leverage and explains part of the cross-country variability. Results remain significant in different subsamples. Consistently with the theory of debt tax shield, the positive correlation between debt and taxation is stronger for highly profitable firms. These findings are robust to different estimation methods and to the use of different proxies for the degree of financial development and the characteristics of the legal system of the country where firms are located. A review of the role played by taxation and institutional characteristics, along with the effect of other firm variables on leverage is presented in section 2. In section 3 we describe our data sources and our variables. Our methodology of estimation and the results of the empirical analysis are presented in section 4. The findings are assessed in section 5. In section 6 we draw the main conclusions. 2. The determinants of leverage In this section we review the main variables that are related to leverage according to corporate finance theory. Firstly, we illustrate some proxies of debt and non-debt tax shield, since taxation is the focus of our paper. Then we review the role played by institutional features, grouped in legal and financial variables; they are relevant because they affect agency cost and information asymmetries. Finally, we describe the relation between firm-level characteristics and leverage. 2.1 Debt and non-debt tax shield In corporate finance literature, the influence of taxation on financial structure is related to the possibility that firms can deduct some costs: interest paid, thus obtaining a debt tax shield; 3

4 depreciation and similar expenses, which represent a non-debt tax shield (DeAngelo and Masulis, 1980). In particular, debt seems positively associated to corporate marginal tax rate, to the ratio of personal taxation on equity income over interest income, and negatively linked to non-debt tax shield (NDTS), to already existing interest rate deductions, and to the probability of future losses (Graham, 2006). There are several variables which measure debt tax shield and each of them have some limitations. 4 A simple measure is the effective tax paid by firms, which are grouped by Nicodème (2001) in two main categories. The backward-looking method measures the weight of tax over past corporate income and takes into account the effects of the business cycle; the forward-looking approach is based on simulated corporate income and considers all taxation rules (King and Fullerton, 1984; Devereux and Griffith, 1998). An issue related to these measures is the endogeneity of taxation status: the more a firm uses the debt tax shield the higher is the probability it incurs in losses, thus reducing its tax rate. A possible solution to this problem is a forward-looking estimation of tax rate before interest deduction (Graham et al., 1998; Alworth and Arachi, 2001). The relevance of negative profits is due to the related tax credits, which implies that marginal tax rate for a firm in the case of a loss is zero. A direct estimation of the marginal tax rate allows to consider tax carrybacks and carryforwards. The problems of this method are that it requires a long time series of firm data and the forecast of future income. However some analysis developed in Graham (1996) show that a simpler approach is possible. Some proxies are the corporate statutory tax rate, a dichotomous or a trichotomous variable. 5 An alternative approach takes into account investor choice between an investment in equity or bond. In a classical tax system dividends are double taxed at corporate and personal tax rate, while interests are only taxed at personal rate. Miller (1977) says that this tax advantage of debt is balanced from the higher interest rate that a firm has to pay to investors who look at net return on bond or equity. In the empirical literature, Miller index is largely used in order to take into account different country tax rates, including the corporate and the personal taxation on dividend and interest payments. 6 However, such synthetic indicator has some shortcomings. First, it depends on which tax rate is applied, as some investors are tax free, while others pay the marginal rate. Second, it is based on the idea that there is an investor which chooses among the subscription of a bond or a share. In a way, it excludes the possibility of loans, which are the main source of finance for firms in European countries and in the US. The same reasoning can be extended to other financial intermediaries: pension funds or insurances could prefer to invest mainly in bonds. Moreover, the increase of capital through retained earnings is ruled out, though cash flow is a very important source of funding for firms. Alternatively, the Miller index implies that retained earnings are discounted in capital gains and that the taxation on capital gains and dividends is the same. 7 4 An alternative way to measure the impact of taxation on debt is through the effects of the financial reforms. For some analysis relative to the Italian case see Staderini (2001), Maurizi and Monacelli (2002), Ziliotti and Benedetti (2007), De Mitri et al. (2008), and Crespi et al. (2009). 5 The dichotomous variable is equal to the statutory tax rate if the taxable income is positive and zero otherwise. A less predictive variable is a dummy equal to zero if the firm has net operating loss in any period or equal to the statutory tax rate otherwise. The trichotomous variable is equal to: a) 0 if both the taxable income is negative and there are net operating losses in any period of the considered time span; b) ½ of the statutory tax rate if the taxable income is negative or there are net operating losses in any period; c) the statutory tax rate otherwise. 6 Monacelli et al. (2001) develop a model with a synthetic tax rate including all these rates. 7 To evaluate these shortcomings it is useful to derive the arbitrage rule from which Miller index is obtained, but excluding some simplifying hypotheses. We consider the three possibilities a firm has to finance its assets: debt, equity or internal funds. Suppose a firm distributes a fraction p of its profits while (1-p) are reinvested. Then the following equivalences must hold: d p( 1 τ c )(1 ατ d ) π = p(1 τ i ) i (1) nd ( 1 p)(1 τ )(1 βτ ) π = (1 p)(1 τ i (2) c cg i ) 4

5 Another proxy of advantages of taxation generally used in the literature is the non-debt tax shield. The drawback of this measure is that it could be positively related to profitability and investment, so that if a high-tax rate and profitable firm invests more and also borrows more, there could be a positive relation between NDTS and debt. MacKie-Mason (1990) addresses this problem interacting NDTS with an ad hoc near tax exhaustion variable and finds a negative relation. 2.2 Legal and financial system The legal system is one of the main factors which reduce conflicts of interests among different firm agents: managers, employees, shareholders and external investors. La Porta et al. (1998) report that the rules of law and their enforcement play a relevant role in financing decisions: if the legal system is not developed and the enforcement of rules is not strict, it is more likely a higher use of contracts that reduce discretion of managers or majority shareholders, such as short-term debt. Bankruptcy law is a specific component of the legal rules that affects more directly the relationships between a firm and its creditors in case of financial distress (Djankow et al., 2008). In fact, there is a trade-off between the preservation of the on going business of profitable firms from their managers and the protection of creditors, proxied by a fast exercise of their rights which in turn incentives managers and shareholders to avoid bankruptcy. As a consequence, the stronger the protection of creditors, ensured by a faster bankruptcy process, the higher could be the leverage. Another relevant variable is the level of corruption of a country, which is connected to the formal application of laws by the judiciary system (Djankow et al., 2003). The possibility of expropriation of external financers from managers or public officials implies that a higher level of short-term debt is expected the more a country is perceived as corrupt. Other variables used in the previous empirical literature are related to the role of financial markets. The main distinction along this line is between bank-based and market-based financial systems. The variables usually considered are bank loans to private sector and the capitalisation of bond or equity market as a percentage of GDP. The idea is that the presence of a more extended banking system should be more related to higher and short-term debt, because in bank-based system loans are the main source of finance for firms, apart from internal funding or shareholder resources. Also a stricter relationship between firms and banks - which could be approximated by the number of banks a firm typically use - would reduce agency problems, thus increasing debt (e. g. Petersen and Rajan, 1994). A bigger development of the bond market could represent an alternative to the banking system and should allow more long-term debt. Leverage is also influenced by the development of the equity market, which facilitates the issue of shares. It should be noted that bank or market oriented financial systems are not unambiguously related to leverage. In fact bank loans could be more expensive than market alternatives because interest rates may cover bank monitoring costs (Diamond, 1991). Hence it is not necessarily the case that leverage is higher in bank-based financial systems. d nd in which the tax rates are τ and c τ (statutory corporate tax rate on distributed and non distributed earnings), c τ i (personal tax rate on interest), τ (marginal personal tax rate on dividend), and d τ (marginal personal tax rate on cg capital gains); i is the return on debt, π is the return on equity; α(β) is the percentage of dividends (capital gains) subject to taxation. The sum of (1) and (2) yields: nd i p + (1 p)(1 τ c )(1 βτ cg ) (3) = π p(1 τ i ) + (1 p)(1 τ i ) nd (1 τ c )(1 ατ d ) d nd The Miller index implies that α =1 and p=1 (or that τ c = τ, c τ d = τ, and cg α = β ). 5

6 2.3 Firm characteristics The importance of firm characteristics and their relation with leverage vary according to the different views of corporate finance theories. The survey of Harris and Raviv (1991) shows that leverage is positively related to firm tangible assets, NDTS, growth opportunities, and size, while it is negatively related to volatility, bankruptcy probability, intangible assets, profitability, and uniqueness of production. In a more recent paper, Murray and Vidhan (2009) find that leverage calculated at market value is positively associated to industry leverage, tangible assets, and size, while it is negatively linked to growth and profitability. However, the impact of size and growth is no longer significant when leverage is calculated at book value. More in detail, profitability could have an ambiguous relation with leverage. On the one hand, pecking order theory implies that the relation is negative, because firms prefer internal funds to debt and equity. On the other hand, trade-off theory suggests that debt is preferred to equity in order to benefit from debt tax shield, which is higher for profitable firms. Also the free cash flow problem suggests a positive relation. In empirical literature a negative relation is usually found. This result could be influenced by bank-firm relationship, which reduces asymmetric information. Tangible assets allow to reduce the agency costs of debt because they can be used as a collateral for financing, especially for loans. A possible substitute for collateral is the presence of stronger banking relationships, so that in this case tangibility should matter less. The variable is anyway positively related to leverage and long-term debt. Size is a widely used proxy for risk, because bigger firms tend to be more diversified and fixed costs associated with their bankruptcy are lower, so that the relation with leverage should be positive. However the sign is not unambiguous because the asymmetric information between insiders and financial markets is lower for bigger firms, which have easier access to external capital, especially if financial markets are more developed. This effect suggests a negative relation with leverage. Age can have an ambiguous relation with leverage. According to the pecking order theory, as a firm becomes older it can rely more on retained earnings to finance its investments; hence a positive relation is expected. However, age could also be related to risk because older firms have a lower incentive to invest in risky projects. This reputation effect could increase leverage because older firms are considered safer from creditors. Intangible assets may represent a measure of opaqueness of a firm, a possible source of funds diversion, and their quality as a collateral is low. All these reasons suggest a negative relation with leverage, especially if protection of creditor rights is low. Growth opportunity can be considered as a form of intangible asset. If asymmetric information is relevant there could be underinvestment problem because debt holders fear that a part of their return could be taken away by shareholders. The consequence should be a negative correlation with leverage. Volatility of earnings is a proxy of risk of bankruptcy and is negatively related to leverage and long-term debt. The effect should be higher if protection of creditor rights is lower. Bankruptcy probability is another negative determinant of leverage, because debt holders would finance less a riskier firm. 3. Data This section summarises the main characteristics of the Amadeus database, which represent the source of firm data. We then illustrate our variables of interest and provide some descriptive statistics. 6

7 3.1 The Amadeus database We use an unbalanced panel data of around 487,000 firm-year observations in 13 European countries during the time span. 8 Amadeus database by Bureau Van Djck is the source of individual data of unconsolidated firm balance sheets and income statements. Most of the sample is represented by unlisted firms (listed firms are only 2 per cent). Data includes the main components of assets and liabilities. Unfortunately, it is not possible to differentiate loans from bonds in order to evaluate the role played by banks. Also it is not possible to distinguish provisions from long-term financial debt. Data of income statement include the main elements; there are few countries for which value added and its components are not reported. We use information on number of employees, turnover, and total asset to classify firm dimension in a way similar to European Commission (EC) definition: small, medium, and large (table 1). 9 We classify firms in four sectors - energy, manufacturing, construction, services - according to their SIC code (table 2). The breakdown of firms among countries by size and sectors is reported in table 3a. It is important to stress some caveats concerning the use of Amadeus data; the different coverage in terms of countries and firm size have some consequences on the representativeness of the sample. In general smaller firms are underrepresented, because there are some minimal requirements to enter into Amadeus database. 10 The ratio of medium and large firms to the total number as reported in Eurostat (2009) is relatively high (table 3b). It is around or above 45 per cent in all the countries we include, apart from Austria and Germany, for which the number of firms is pretty low; the comparison is not possible for Belgium, Greece and Ireland. De Socio (2010) presents a descriptive analysis of the database and of the main indexes derived from balance sheets and income statements, and develops an econometric analysis of the persistence of cross-country differences when firm sector and dimension are considered. The countries taken into account are those that participated to Eurozone in 2002 and the United Kingdom. This allows to consider countries with the same currency and policy interest rate, but the United Kingdom. The time span considered is restricted to the period for three main reasons. First because there is a substantial increase in the number of firms included with respect to previous years. In this way, the results are less influenced by the changes in the sample. Second, because in these years the cross-country differences in firm leverage remain similar and this allows to evaluate the role played by country-level variables, with particular focus on taxation. Finally, in this time span the financial structure of firms is not influenced by the changes in the economic and financial environment following the financial crisis and the recession. 3.2 Variables description Our variable of interest - denoted Leverage - is defined as the ratio of financial debt (including provisions) over the sum of financial debt and equity, both at book value. Summary statistics are reported in table 4, which shows the high variability of firm variables within and between countries. The mean value of leverage in the sample period ranges from 35 per cent in France to 54 per cent in Italy, which has the highest value among all the countries we analyze. Other highly levered firms are in Germany and Greece (53 per cent and 49 per cent, respectively), while in Finland, Luxembourg and Netherlands leverage is below 40 per cent. 8 We restrict the sample to around 373,000 observations in our regression analysis because of outliers in our variables of interest. 9 See for the EC definition. We denote a firm that is micro or small according to EC as a small firm. 10 Smaller firm correspond to more then 98 per cent of firms in every country of European Union (Eurostat, 2009), while the firms included in Amadeus have either total assets greater than 20 million euros, or turnover greater than 10 million euros, or more than 150 employees. 7

8 The source of corporate statutory tax rate (Tax) is European Commission (2010). Between 1994 and 2004, the tax rate decreased among the countries analyzed in this paper, while it remained relatively stable in the period we consider. We rely on statutory corporate tax rate as our taxation variable, since it is a reliable measure which captures cross-country differences. Moreover, it avoids some of the issues discussed in section 2.1: the endogeneity related to the effective tax rate, the estimation of future income to obtain the marginal tax rate, and the limitations and measurement problem of personal taxation of Miller index. We consider also institutional variables and firm characteristics, to control for their effect on leverage summarized in section 2. As regards the institutional indicators, we consider some proxies from Doing Business of the World Bank: a) Investor protection (Invprot), an index (0-10) which summarises the strength of the control over managers acts and measures of the protection offered primarily to shareholders but also to debt holders; 11 b) Legal right protection (Legright), an index (0-10) which measures how much laws protect creditor rights; c) Easiness to close a business (Closbus), which is the number of years required to close a business and a proxy of bankruptcy easiness. We include data from Transparency International, which provides the Corruption Perception Index (Cor), a measure of how public officials use their power for private gain. The index goes from 0 to 10 and a higher value indicates a lower level of corruption. Also we use data on bonds and shares of non financial corporations (NFC), taken from financial accounts, whose source is Eurostat; data on loans from the banking sector are taken from European Central Bank and Bank of England. 12 As measures of the development of financial markets, we calculate the ratio of bonds to financial debt or GDP (Bond or Bond_gdp, respectively), the ratio of bank loans to financial debt or GDP (Loan and Loan_gdp, respectively), and the ratio of total shares or listed shares to GDP (Sh_gdp and Qsh_gdp, respectively). We prefer these ratios to others used in literature, like stock market capitalisation or loans to the private sector to GDP, because our measures include only firms data and are not influenced by listed shares of financial corporations or loans to households. It should be stressed that these country-level data are valuated at market price, so they are different from book value firm-level data. We also consider a proxy of the strength of bank-firm relationship: the median number of banks that firms typically use in each country (Numba), taken from Ongena and Smith (2000). All the country-level variables are summarized in table 5, which reports annual averages in the sample period. The highest tax rates are in Germany (38.6 per cent) and Italy (37.3 per cent), which are also two of the countries with the highest leverage. The other countries have rates between 27 per cent and 35 per cent, except for Ireland (12.5 per cent). High levered firms are more frequent in countries with a higher level of perceived corruption, more bank loans, and lower value of total shares over GDP; these countries also rank lower in terms of legal rights (except for Germany). Low levered firms are more typical in countries with a lower rate of taxation, a lesser level of corruption, and more developed financial markets. The United Kingdom has a standing alone position because its firms are quite high levered but it shares common characteristics with low levered countries such as more developed financial markets and a stronger investor protection. Correlations between country level variables are reported in table 6. We use these correlation to select the variables for our regression analysis presented in section 4. There is a high correlation between the corruption index and measures of financial markets development and legal system. A stronger protection of legal rights, a higher development of financial market, and a smaller banking sector are associated with lower perceived corruption. Given this evidence, we drop the corruption variable from our base regression analysis to avoid multicollinearity problem. A similar reasoning has been applied to the number of banks, which is highly correlated with legal variables and taxation. We used these two excluded variables in robustness checks presented in section This variable is an extension of the dummy which identifies civil law countries. In our sample they are UK and Ireland, which have by far the highest values. 12 The use of these variables does not allow to consider Luxemburg, for which these statistics are not available. 8

9 Finally, as firm variables we calculate from Amadeus data: i) ROA (return on assets) calculated as operating income (OI) over total assets; ii) Tangibles defined as tangible fixed assets over total assets; iii) Growth equal to the annual percentage variation of total assets; iv) Intangibles computed as intangible fixed assets over total assets; v) NTDS defined as depreciation over fixed assets; vi) Sd_ROA calculated in each year and country as the standard deviation of ROA in the sector and the size class in which the firm is classified; vii) Z_score computed as a variant of Altman s (1968) indicator of bankruptcy similarly to Graham (1999). Our indicator is: Z 1.2* working _ capital + 1.4* retained _ earnings + 3.3* OI * sales = total _ asset viii) NWC (net working capital) given by the ratio between commercial credit plus inventories minus commercial debt over total assets. We introduce this variable because we focus on financial debt and our leverage measure does not include trade debt, which mainly depends on commercial relations and not on the effect of debt tax shield. Therefore, we control for the role of commercial net assets, which could have a positive effect on leverage because trade credit and inventories are usually financed by commercial debt and short term (banking) debt. A positive relation is hence expected with leverage and short term debt. We also use information about firm age (Age) and whether it is listed or not (Listed). 4. Regression analysis and main results This section illustrates our model for the level of leverage and estimation methods. Then we summarise the main results for different estimation methods and various subsamples. 4.1 The model We employ a general static panel data model that can be written as: y = x β + z γ + α + u ikt ' kt ' it ik ikt for i=1,, I; k=1,, K; t=1,, T i, (1) where y ikt is the leverage of firm i in country k at time t, β and γ are the parameter vectors for the corresponding country variables x kt and firm variables z it, respectively; we also include time dummies; α ik is the unobserved firm-specific heterogeneity and u ikt is the stochastic error, which is allowed to be cross-firm heteroskedastic and serially autocorrelated. 13 The unobserved heterogeneity may capture all time-constant effects either referred to firm characteristics, as the skill of managers, or connected to omitted country variables. These effects may give inconsistent estimation of the parameters of interest if they are related to the observables x kt and z it. We discuss this issue in the following section. We use a static model of leverage because we do not consider an economic model of the dynamic adjustment towards an optimal level of leverage. In this sense a dynamic model which includes the first lag of leverage could be misspecified and the results could depend on the instrumental variables required to perform the estimation. Moreover, we have a short time span and 13 We tackle this issue by clustering at the firm level using Huber-White sandwich variance estimator which is consistent when the errors are heteroskedatic or serially correlated over the panel observations. 9

10 our country level variables have limited time-variability, so the use of a standard GMM approach for dynamic model (Arellano and Bond, 1991) would imply a loss of information about the variables we are more interested in. As country variables we consider: a) Tax given by the corporate statutory tax rate; b) Sh_gdp as the market value of non financial corporations; c) Qsh_gdp as financial market development; d) Invprot as a proxy of legal characteristics. The firm variables we include in our base regression are the most relevant in corporate finance empirical literature (see section 2.3): profitability, tangible assets and age. We control for firm size and sector using the dummies described in section 3.1. As explained before, we also introduce a measure of net working capital The estimation methods We firstly estimate our model with a pooled ordinary least square (OLS). The method yields consistent estimates for our variable of interest (Tax) if there is no correlation between taxation and the unobserved heterogeneity. This hypothesis may be reasonable as we control for many institutional variables, which may capture most of the omitted country variables effect. Also we assess the impact of the inclusion of different country-level variables in Section 5. We apply a second method in order to take into account the possible relation between taxation and some unobserved variables at country level: the correlated random effect (CRE) approach as proposed by Wooldridge (2010). The method is an extension of Mundlak (1978) for an unbalanced panel and relies on the fact that it is always possible to write the linear projection of the individual unobserved effects as: α ik ϕ = + w ' ikξ + a ik (2) where w ik is the time mean of the time-variant covariates in w ikt = [x' kt z' it ]', including the time dummies as we have an unbalanced panel, and a is the stochastic error which has zero mean and is uncorrelated with w ik, by assumption. The combination of (2) and (1) yields: ' ' ' y ikt = xktβ + zit γ + + wikξ + ai + k ikt i k ϕ u (3) It is possible to show that the estimated coefficient of w ikt is equivalent to the fixed effect estimator of the main equation (1). The estimated parameters are consistent under the standard hypothesis of the within estimator, provided that there is enough time variability. Moreover, the method gives us the possibility to directly test the hypothesis of no relation between the unobserved effects and the time-variant covariates (regression-based Housman test), looking at the coefficients of w ik. We finally consider the impact of firms heterogeneity in the estimated relation between leverage and determinants of capital structure. The underlying idea is that the impact of the regressors could differ according to the firm position in the leverage distribution. So we run quantile regressions for the 25th -, 50th - and 75th-percentile. Quantile regressions may give informative results by modelling non-linear effects (Koenker and Basset, 1978) and are robust to heavy tailed regression errors. 14 The regressions exclude the outliers of all firm variables. We exclude values below the 1st-percentile and above the 99th-percentile. We calculate the outliers of the variables for each country and each year. 10

11 4.3 Main results: taxation We present our main results in tables 7-9. In table 7 we compare the results of different estimation methods. We begin with pooled OLS, in a base form (1) and in a modified version (2) where country dummies replace country-level variables. This second estimation allows to test if there are major changes in firm-level coefficients due to the introduction of different country-level variables. The fact that firm-level estimates are basically unchanged and the fit is quite identical suggests that taxation and institutional variables can be used instead of country dummies to evaluate cross-country differences. Then we present the results of the CRE (3) and of the random effect (4) estimators. We finally show the quantile regression results (5-7). In all estimations the influence of taxation remains positive and significant. In general, the OLS estimator looks like an upper bound while the CRE estimation like a lower bound. The estimated value of the OLS regression indicates that if tax rises by 4 percentage points - which corresponds to a standard deviation - then leverage goes up, on average, by 3 percentage points. The impact is also relevant when the cross-country effect is measured considering variation between the lowest and the highest tax rate (Ireland, 12.5 per cent and Germany, 38.6 per cent): it amounts to a difference in leverage of nearly 19 percentage points. The influence of taxation on leverage is consistent with the evidence found by Bartholdy and Mateus (2008) and Pfermayer et al. (2008), even if their estimates are restricted to a sample of manufacturing firms. Our result is somehow lower, because we are also including proxies of the financial markets and the legal system and not all of them are considered in these previous works. The effect of taxation as results from the CRE, although lower, is still economically significant: if tax rises by 4 percentage points then leverage goes up, on average, by 1 percentage points. The regression-based Housman test, which is read on the coefficient of the time mean of taxation, says that we cannot neglect the relation between the unobserved heterogeneity and taxation. Nevertheless, these results should be read with caution given the rather small time variation in this variable. Quantile regressions confirm that the heterogeneity of leverage affects the coefficients of the independent variables. As regards taxation, it has higher impact if firms have lower leverage: the estimated coefficient decreases from 1 for the first quartile to 0.6 for the third quartile. These results are coherent with the expectation that more levered firms are already using debt tax shield and have a lower incentive to increase their debt, also because they have a higher cost of distress. Tables 8-9 report estimates for different subsamples using OLS and CRE, respectively: unlisted firms, since they could have different relations with the variables we are considering; large firms, which could be less opaque and with more reliable data; manufacturing companies because they are a more homogeneous group and their results are comparable with previous studies; highly profitable firms, 15 whose marginal tax rates should be around the statutory tax rates. OLS estimates are significantly positive in all subsamples and with a higher coefficient, around 0.8. We find this result for larger and manufacturing firms. As expected, the impact is bigger on highly profitable firms, for which tax gain from increased leverage should be higher because they are more likely to obtain a positive income. CRE estimates confirm a lower but significantly positive effect of taxation on leverage in all subsamples. 4.4 Main results: institutional and firm variables OLS regression provides evidence of a significant relation between institutional variables and leverage (Table 7, col. 1; Table 8 for subsamples). The development of financial markets (Qsh_gdp) has a negative effect on leverage, which is an expected result. Also the impact of the market value of non financial corporations (Sh_gdp) is strongly negatively related with leverage, which implies 15 Highly profitable firms are defined each year t as those firms with positive operating income in all sample years and positive profit in the year t. 11

12 that a higher market value of the listed and unlisted companies of a country is related to a lower leverage. The economic effect of Qsh_gdp is somehow smaller than Sh_gdp: an increase of a standard deviation is related to a decrease of 2.6 and 4.4 percentage points in leverage, respectively. The magnitude is somehow bigger when cross-country effect is measured comparing the countries with the highest and the lowest values. It amounts to 7.6 percentage points for Qsh_gdp (which goes from 28 per cent in Italy to 105 per cent in Finland) and to 10.4 percentage points for Sh_gdp (which ranges from 79 per cent in Greece to 218 per cent in Belgium). Finally, the influence of the investor protection is positive and also economically significant: an increase of a standard deviation also raises leverage by 3.4 percentage points. The impact in terms of cross-country differences is much bigger and equal to 14.4 percentage points as the variable ranges between 3 in Greece and 8.3 in Ireland. As regards the sign and the significance of the individual characteristics, they are robust in different specifications and in various subsamples. We evaluate the effects on leverage as a change of a standard deviation of the independent variable. Profitability has the biggest impact on leverage, confirming that more profitable firms are less levered, presumably because they have more internal resources to finance investment. If profitability decreases by 11 percentage points, leverage rises by 5 percentage points. Both tangible assets and net working capital show a positive relation with leverage and their effects are similar: for both variables an increase of around 24 percentage points raises leverage by more then 4 percentage points. Age has a negative effect (around -2 percentage points if a firm is 20 years older), presumably because older firms produce more cash flow to finance their investments. Listed firms show a lower leverage (around 7 percentage points) mainly due to their easier access to alternative financing markets. Finally, there is an effect of sector and size dummies, even if it is not always statistically significant. In general, leverage is higher for construction firms and lower for manufacturing ones, while large companies are more levered than smaller ones. 16 CRE estimates (Table 7, col. 3; Table 9 for subsamples) confirm OLS results except for the Qsh_gdp variable, which shows a positive sign. As regards to the individual variables, the CRE estimator presents lower values. These estimations should be preferred to the pooled OLS, because it is likely that firm unobserved heterogeneity may be correlated with firm-level variables. Also individual variables have higher time variability, which is an important feature when only within variation is used in the estimation. 5. Robustness checks We test our OLS results in several ways and along three main dimensions: 1) restriction of estimations to subsamples; 2) use of different proxies of institutional variables; 3) introduction of further firm-level characteristics. We finally assess if our results are influenced by endogeneity issues or by how standard errors are clustered. First we check the robustness of our results with respect to the exclusion of each country, leaving it out one by one and our estimated coefficients remain all significant. Results are also stable when we run the regression for each single year. Then we focus our attention on firm sector or size by running a regression for each category. The impact of taxation on leverage is confirmed in all class sizes and is lower for small firms. It remains also significant for the different sectors, except for the energy one, and is greater in the construction sector. As a second control of our results we use different country-level variables. We evaluate the effects of the inclusion of different combinations of legal system variables (Legright or Closbus instead of Invprot) and financial development proxies (Bond, Bond_gdp, Loan, or Loan_gdp instead 16 We also considered a continuous definition of firm dimension using the logarithm of total assets. We find a positive relation, but the results are not always statistically significant. 12

13 of Qsh_gdp). 17 We also test the effect of the exclusion of Sh_gdp, which we introduced in our paper to take into account the market value of all non financial corporations in a country. In general the relation between taxation and leverage remains positive and strongly statistically significant; the use of different proxies of legal system reduces the coefficient of taxation to a range between 0.5 and 0.6; our base regression includes Invprot because it is the most significant variable and it is related to the concept of civil law countries. The use of the other proxies of financial market development increases the estimated parameter of taxation with respect to our base regression to a range between 0.8 and 0.9. Finally the elimination from the base model of Sh_gdp only slightly reduces the impact of taxation. 18 We also try to include in our base regression all the country-level variables we consider using principal component analysis (PCA) in two different ways. We initially derive two new variables from the six proxies of financial development (mkt) and from the three variables of legal system (law) separately by taking the first principal component. 19 We include these two variables in the base regression in place of Qsh_gdp and Invprot and the results on the other variables are unchanged. These two variables are significant and with the expected sign: negative for mkt and positive for law. However, since they are somehow correlated (0.5), we also derive a single proxy from all the nine variables (mkt_law) and we use it in our base regression. 20 The results are nearly unchanged. Some empirical studies suggest that corruption is associated with higher leverage (e.g. Fan et al., 2011). We also include in our regression Cor, excluding Qsh_gdp and Invprot to avoid collinearity. When perceived corruption is introduced, the result of taxation is somehow smaller (0.5), but still strongly significant. We find a high correlation between mkt_law and Cor (0.8), which suggest that the Perceived Corruption Index may sum up both the financial development and the legal structure of countries. 21 A third class of further control concerns the inclusion of other firm-level variables. We checked that they do not change our results of OLS base regression. We introduce asset growth, intangible assets, NDTS, volatility of earnings, and Z-score. Estimation of previous variables are similar in terms of sign and significance; some differences are mainly due to the change in the sample size. The firm-level variables added to the regression are statistically significant, but not always with the expected sign. On the one hand, NDTS and Z-score are negatively linked to leverage, while asset growth has a positive relation; on the other hand, intangible assets and volatility of earnings are positively related to leverage, while the relation suggested from the theory is negative. We do not include these variables in our base regression for a few reasons. First and more important because they reduce the number of observations (except for volatility of earnings), and for some of them we lose nearly all observations for Austria (asset growth) or Ireland (intangible assets). Also their contribution to explain leverage variability (in terms of increasing in adjusted R 2 ) is very low. Moreover, these variables do not seem good proxies of the measures suggested by the theoretical models. 17 We do not consider Numba because it is correlated with other variables. 18 The signs of the variables we introduce are as expected: positive (but not always statistically significant) for Legright and negative for Closbus; negative for Bond and Bond_gdp; the sign of Loan and Loan_gdp depends on the other variables included in the estimation. 19 The first principal component seems a good proxy of all the variables we consider in both cases. It accounts for more then 70 per cent of total variance, the eigenvectors are similar in absolute value for all the variables considered, and its correlation with the base variables is above The first principal component accounts for 60 per cent of the variance and is strongly correlated (above 0.75) with all the base variables, except Closbus. 21 As reported in section 3, the perceived corruption is negatively correlated with financial development and investor protection, which corporate theory indicates as determinant of leverage. However these two country characteristics are in opposite relation with leverage: negative for financial development and positive for investor protection. Since Cor includes these two contrasting characteristics, we prefer to use in our base regression two different variables instead of just this proxy. 13