Corporate Governance, Incentives, and Tax Avoidance

Transcription

1 Corporate Governance, Incentives, and Tax Avoidance Christopher S. Armstrong The Wharton School University of Pennsylvania Jennifer L. Blouin * The Wharton School University of Pennsylvania blouin@wharton.upenn.edu Alan D. Jagolinzer Leeds School of Business University of Colorado jagolinzer@colorado.edu David F. Larcker Graduate School of Business Stanford University larcker_david@gsb.stanford.edu April 2014 Abstract: This paper examines the link between corporate governance, managerial incentives, and tax avoidance. Similar to other investment opportunities, unresolved agency problems may cause managers to over- or under-invest in tax avoidance relative to the preferences of shareholders. Using quantile regression, we find that the impact of corporate governance on tax avoidance is most pronounced in the upper and lower tails of the tax avoidance distribution, but not at the mean or median of this distribution. Specifically, we find a positive relation between the financial sophistication and independence of boards and tax avoidance in the lower tail of the tax avoidance distribution, but a negative relation in the upper tail of the tax avoidance distribution. However, we find no relation between corporate governance and tax avoidance at either the conditional mean or median of the tax avoidance distribution. These results suggest that corporate governance tends to decrease extremely high levels of tax avoidance and increase extremely low levels of tax avoidance, which may be symptomatic of over- and under-investment, respectively, by managers. Our results also suggest that inferences about these relations that are drawn from the conditional mean and median and unlikely to be representative across the entire tax avoidance distribution. JEL: G34, H25, H26, K34, M41 Keywords: Tax aggressiveness; FIN 48; tax avoidance; CEO incentives; corporate governance *Corresponding author We thank Shane Heitzman and workshop participants at University of Delaware, ESADE, Goethe University, University of Iowa, University of Oregon, University of Toronto research conference, Texas A&M University, and Washington University for helpful comments. Jagolinzer thanks the EKS&H Faculty Fellowship for financial support.

2 1. Introduction There has been a recent surge in research that seeks to understand the sources of variation in tax avoidance (e.g., Shevlin and Shackelford, 2001; Shevlin, 2007; Hanlon and Heitzman, 2010). The benefits of tax avoidance can be economically large (e.g., Scholes et al., 2009) and tax avoidance can be a relatively inexpensive source of financing (e.g., Armstrong et al., 2012). However, aggressive tax avoidance may be accompanied by substantial observable (e.g., fines and legal fees) and unobservable (e.g., excess risk and loss of corporate reputation) costs. Although understanding the factors that influence managers tax avoidance decisions is an important research question that has broad public policy implications, relatively little is known about why some firms appear to be more tax aggressive than others. We examine whether variation in firms corporate governance mechanisms explains differences in their level of tax avoidance. We view tax avoidance as one of many investment opportunities that is available to managers. Similar to other investment decisions, managers have personal incentives to engage in a certain amount of tax avoidance that may not be in the best interest of shareholders, thereby giving rise to an agency problem. From the perspective of the firm s shareholders, unresolved agency problems with respect to tax avoidance can manifest as either too little or too much tax avoidance. As with other agency problems, certain corporate governance mechanisms can mitigate agency problems with respect to tax avoidance. Few papers directly examine the link between corporate governance and tax avoidance. Minnick and Noga (2010) investigate whether several measures of corporate governance are associated with a variety of proxies intended to capture firms level of tax avoidance, but find little evidence that governance is associated with avoidance. Desai and Dharmapala (2006) report the surprising result that firms that are poorly governed but where managers have high - 1 -

3 levels of equity incentives engage in less tax avoidance. They interpret this result as evidence that tax avoidance and managerial rent extraction are complementary activities, which implies that the level of a firm s tax avoidance is increasing in the strength of its corporate governance. Rego and Wilson (2012) find that firms at which managers have high risk-taking equity incentives engage in more tax avoidance. However, they fail to find any evidence that governance mechanisms other than executives equity incentives affect this relation. In a concurrent paper, Robinson et al. (2012) examine the association between tax avoidance and audit committee financial expertise. They report evidence that audit committee financial expertise is generally positively associated with tax planning, but that this association is negative where they deem tax planning to be risky (i.e., aggressive). Overall, the few papers that examine the relationships between corporate governance, managerial equity incentives, and tax avoidance have produced results that are quite mixed. One common theme across prior studies is that their inferences are based on estimates of the conditional mean of the tax avoidance distribution. However, estimates of the conditional mean may not be representative of the relation between governance and tax avoidance at other parts of the tax avoidance distribution. Rather than use traditional econometric methods that focus on either the conditional mean or median of the relation between tax avoidance and corporate governance, we utilize quantile regression to assess this relation across the entire tax avoidance distribution. This research design follows naturally from our conjecture that corporate governance should have a differential impact on extreme levels of tax avoidance. For example, it is possible that boards that better understand the net benefits from tax strategies would encourage more tax planning at lower levels of the tax avoidance distribution because this improves cash flows with little accompanying risk. In contrast, boards might discourage additional tax - 2 -

4 avoidance at higher levels of the tax avoidance distribution because high tax avoidance may impose costs on the firm (e.g., regulatory or reputation) that exceed the marginal benefits of additional tax savings. We examine a comprehensive sample of firms between 2007 and 2011 and find that CEOs risk-taking equity incentives exhibit a positive relationship with the level of tax avoidance. More importantly, we find that this relationship is stronger in the upper tail of the tax avoidance distribution. This result is consistent with managerial risk-taking incentives being an important determinant of aggressive tax choices that are likely to entail more risk. We also assess the impact of other governance mechanisms on firms tax avoidance. In particular, we examine attributes of the board of directors, including representation by financial experts and the independence of directors, as measures of the awareness of the net benefits of investment in tax avoidance and the ability to monitor managers tax avoidance decisions. We find that the relationship between board financial expertise/independence and firms level of tax avoidance varies substantially across the (conditional) tax avoidance distribution. Specifically, we observe a positive relation in the lower tail of the tax avoidance distribution, which is likely to be symptomatic of under-investment in tax avoidance. In contrast, we observe that this relation is negative in the upper tail of the tax avoidance distribution, which is likely to be symptomatic of over-investment in tax avoidance. Together, these findings suggest that more financially sophisticated and independent boards recognize the potential agency problems that would otherwise give rise to extreme levels of tax avoidance and, in these cases, constrain managers tax avoidance decisions. Consistent with the results in Desai and Dharmapala (2006), we do not find a relation between tax avoidance and an interaction between their indicator for good governance and a - 3 -

5 measure of top executives stock option compensation using ordinary least squares (OLS) estimates of the conditional mean. However, quantile regression estimates indicate that there is a negative relation between tax avoidance and the interaction between good governance and stock option compensation in the upper tail of the tax avoidance distribution and no relation in the lower tail. These estimates suggest that the interaction between executive stock option compensation and good corporate governance mitigates over-investment in high levels of tax avoidance. Thus, in contrast to Desai and Dharmapala (2006), we find that corporate governance appears to be related to managers tax avoidance decisions, but only for high levels of tax avoidance. The remainder of the paper consists of six sections. Section two develops our hypotheses related to the relation between corporate governance and tax avoidance, and discusses how the relation may vary for different levels of tax avoidance. Section three describes our sample selection. Section four discusses our research design and explains our choice of quantile regression estimation. Section five presents our primary empirical results and inferences. Section six discusses additional supplemental analysis and reexamines Desai and Dharmapala s (2006) findings. Section seven provides concluding remarks. 2. Prior Literature and Hypothesis Development 2.1 Prior Literature A mature stream of corporate tax research examines the determinants of effective tax rates and book-tax differences (e.g., Gupta and Newberry, 1997). Subsequent studies (e.g., Mills and Newberry, 2001 and Cloyd et al., 1996) focus on the book-tax tradeoffs that are associated with various tax avoidance opportunities (i.e., some tax avoidance strategies reduce both taxable and financial statement income, whereas others affect only taxable income). Although this prior - 4 -

6 research is useful, it provides little insight regarding why some firms seem to avoid taxes using tax planning more than others (Shackelford and Shevlin, 2001). This gap in the literature spurred a series of papers that more directly examine the determinants of tax avoidance. 1 For example, Dyreng et al. (2010) report evidence that executives who were previously employed by firms that are characterized as tax aggressive seem to import this aggressiveness to their new employer. Slemrod (2004), Crocker and Slemrod (2005), and Chen and Chu (2005) suggest that corporate tax noncompliance (i.e., extreme tax avoidance) could result from the design of managers incentive-compensation plans. Consistent with this notion, there is empirical evidence that tax avoidance is associated with greater levels of incentive compensation (e.g., Phillips, 2003; Armstrong et al., 2011; Rego and Wilson, 2012). There is little research that directly examines whether (or how) corporate governance affects tax avoidance. Desai and Dharmapala (2006) develop and subsequently test a model that links managers equity-based compensation to aggressive tax avoidance. They conjecture that there are complementarities between tax-sheltering and rent extraction because well-governed firms are assumed to provide managers with greater incentives for tax avoidance because the presence of other governance mechanisms will prevent managers from extracting the rents generated by their tax avoidance activities. However, poorly governed firms will not provide incentives for aggressive tax avoidance because the lack of monitoring from their governance mechanisms would otherwise allow managers to extract the rents generated by their aggressive tax planning. 1 A related set of research attempts to identify and validate measures of tax avoidance. For example, Frank et al. (2009), Wilson (2009), Lisowsky (2010), and Dyreng et al. (2008) develop alternative measures of tax avoidance. Since the passage of the Financial Accounting Standards Board (FASB) s Interpretation No. 48 (Accounting for Uncertainty in Income Taxes) (hereafter FIN 48), several recent papers measure firms tax aggressiveness using the magnitude of their uncertain tax benefits (e.g., Rego and Wilson, 2012; Lisowsky et al., 2012). As we discuss in more detail below, we carefully consider the specific attribute of tax avoidance that is called for by our research question (Hanlon and Heitzman, 2010)

7 Recent work questions several of the fundamental assumptions that underlie Desai and Dharmapala s model. For example, their model assumes that managers can extract rents generated by tax avoidance because operational complexity (and the accompanying information asymmetry) is a requisite condition for tax avoidance. However, the precise channels through which managers extract rents derived from tax avoidance is not clear and there is limited empirical evidence that managers do actually extract rents generated by tax avoidance. 2 Desai and Dharmapala (2006) also implicitly assume that equity-based compensation does not mechanically create tax shields. However, Seidman and Stomberg (2011) directly challenge this assumption and report that firms with higher levels of equity compensation are less likely to require (i.e., benefit from) additional tax shields from tax avoidance. Seidman and Stromberg (2011) explain that Desai and Dharmapala s association between equity compensation and tax avoidance can be explained by tax exhaustion. 3 Finally, Desai and Dharmapala (2006) implicitly suggest that less rent extraction occurs at firms that are poorly governed. However, this explanation is somewhat counterintuitive if one presumes that managers have more opportunities to extract rents at firms with poor governance and therefore less monitoring oversight. 2 Desai et al. (2007) present evidence that Russian oligarchs appear to extract meaningful rents from firms that avoid more taxes. However, the authors do not find evidence that this is the case for Russian firms that operate in regulated (e.g., U.S.) markets. Blaylock (2011) also fails to find evidence that managers of U.S. firms extract economically meaningful rents through tax avoidance. 3 Graham et al. (2004) suggest that firms require fewer alternative tax shelters if they utilize more stock option grants. This tax exhaustion occurs because option exercises provide material tax deductions, which make it less necessary to engage in alternative tax avoidance strategies. This same reasoning can also be found in the equity incentives literature. For example, Core and Guay (1999, 159) argue that when future corporate tax rates are expected to be higher, the future tax deduction from deferred compensation becomes more favorable relative to the immediate tax deduction received from cash compensation

8 In a related concurrent study, Robinson et al. (2012) also examine the association between incentives and governance for both general and risky tax avoidance. 4 Robinson et al. (2012) report evidence that the proportion of accounting experts on the board is associated with more general tax planning, but less risky tax planning. To the degree their measure of risky tax planning explains variation in the right tail of the tax avoidance distribution, our findings regarding the role of accounting experts on tax avoidance are consistent with those of Robinson et al. (2012). However, one important difference between our study and Robinson et al. is that we also examine the effect of governance and incentives on tax avoidance in the left tail of the tax avoidance distribution. Firms in the left tail of the (conditional) tax avoidance distribution are those that engage in less tax avoidance than expected (i.e., under-shelter ), and constitute an important and unresolved puzzle in the tax literature. 5 Our findings shed light on the relation between various governance mechanisms and tax avoidance for these firms. In summary, although extant literature provides some insight into the role of incentives on tax avoidance, inferences are still limited regarding whether (and how) corporate governance influences firms tax avoidance. Moreover, the evidence that does exist is confined to explaining the conditional mean of the tax avoidance distribution, but does not describe the relation at other parts of the distribution most notably in the extreme tails. It is particularly important to examine these parts of the tax avoidance distribution because there are likely to be substantial economic losses for the shareholders of firms that engage in too little tax planning and 4 Robinson et al. (2012) measure risky tax planning using estimates of tax shelter likelihood derived from the level of activity in tax havens (Dyreng and Lindsay, 2009; Balakrishnan et al., 2012) and the shelter estimation models described in Wilson (2009) and Lisowsky (2010). They also use the predicted uncertain tax benefits as an additional proxy for firms involvement in risky tax planning following Cazier et al. (2009) and Rego and Wilson (2012). 5 The lack of evidence of significant costs of tax avoidance is frequently referred to in the literature as the undersheltering puzzle (e.g., Desai and Dharmapala, 2006; Weisbach, 2002; Hanlon and Heitzman, 2010; and Gallemore et al., 2012)

9 substantial penalties associated with tax planning that is too aggressive (e.g., Hanlon and Slemrod, 2009) Research Hypotheses Managerial incentives If the value of a CEO s equity portfolio is sensitive to changes in stock price and the CEO believes that the expected tax savings both increases stock price and exceeds the expected costs of aggressive tax positions (including any personal costs borne by the manager such as effort and reputational costs), the sensitivity of the CEO s equity portfolio value to changes in stock price (i.e., portfolio delta) should be positively associated with tax avoidance. Alternatively, the sensitivity of a CEO s equity portfolio value to changes in stock price may be negatively associated with tax avoidance if it entails a sufficiently large increase in risk. For example, Armstrong et al. (2012) note that a CEO s equity portfolio delta amplifies the effect of equity risk on the total riskiness of the manager s portfolio, generally discouraging risk-averse managers from taking risky projects. Because of these conflicting effects, it is theoretically ambiguous whether CEOs equity portfolio delta will be positively or negatively associated with tax avoidance across the distribution. If CEOs believe that more aggressive tax avoidance increases stock price volatility, we expect the sensitivity of CEOs equity portfolio value to changes in return volatility (i.e., equity portfolio vega) to be positively associated with tax avoidance. Moreover, the magnitude of this relation should be higher (lower) in the right (left) tail of the tax avoidance distribution where there is more (less) risk to the firm. This prediction is similar to that in Rego and Wilson (2012), except that we also predict a stronger relationship at higher levels (and conversely a weaker relationship at lower levels) of the tax avoidance distribution

10 Board characteristics The function linking shareholder value to tax avoidance is likely to be concave with an interior optimum. In particular, there are likely to be positive net benefits (e.g., cash savings) from engaging in tax avoidance up to a firm-specific optimum. Beyond this point, there may be diminishing marginal net benefits to tax avoidance because of costs related to structuring complicated tax transactions, an inability to repatriate and invest foreign earnings, and potential political, regulatory, or reputational costs that are detrimental to future operations. 6 In contrast, managers can increase their compensation by increasing risk, and they may therefore have an incentive to over-invest in tax avoidance, particularly if they have a shorter horizon than shareholders. These and other differences between managers and shareholders preferences give rise to agency problems with respect to tax avoidance. If shareholders and managers have different preferences for tax avoidance then governance mechanisms may be structured to influence managers tax avoidance decisions. For example, certain governance mechanisms may prevent (or mitigate) over- and under-investment in tax avoidance. Although firms typically utilize a variety of governance mechanisms, we focus on those that we believe are most likely directly associated with tax decisions. Specifically, we focus on the financial sophistication and independence of the board. We expect that more financially sophisticated boards will better understand and monitor managers tax avoidance decisions. This would be consistent with recent guidance that recommends greater awareness in 6 For example, some studies suggest that firms incur potential tax penalties and reputational costs (e.g., Chen et al., 2010) or face significant costs defending aggressive tax positions (e.g., Rego and Wilson, 2012). Although there is little direct evidence supporting the existence of explicit reputational costs (e.g., Gallemore et al., 2012; Hanlon and Slemrod, 2009), there is clearly some underlying cost of extreme tax avoidance because not all firms have extremely low effective tax rates

11 the Boardroom of the importance of tax issues 7 and also recommends that tax issues, including implementing and monitoring tax planning, should be placed on the audit committee s agenda. 8 Accordingly, we predict a positive (negative) relation between the financial sophistication of boards and tax avoidance in the lower (upper) tail of the tax avoidance distribution. Similarly, we expect that more independent boards are more likely to provide monitoring and oversight of managers choices to mitigate agency problems related to tax avoidance (e.g., Desai and Dharmapala, 2006). Thus, we predict a positive (negative) relation between board independence in the lower (upper) tail of the tax avoidance distribution. 3. Sample Selection Our sample selection starts with all firms listed on Compustat for the fiscal years for which we have data to compute at least one of the tax attributes that we define below. We delete foreign registrants and firms designated as real estate investment trusts (REITs) because these firms are subject to different tax rules. We also delete firms that have an annual average stock price of $1.00 or less per share and firms with average total assets of $10,000 or less. These requirements yield 12,275 firm-year observations. We then retain firm-year observations for which we have data available for our control variables (defined below). This yields 7,231 firm-year observations. Finally, we retain firm-years for which we have data available for our governance and incentives variables. 9 This yields a final sample of between 3,137 and 4,128 firm-year observations depending on the measure of tax avoidance. 7 Jeffrey Owens, Good Corporate Governance: the Tax Dimension OECD Forum on Tax Administration, September Deloitte Hot Topics: Taxes: What the audit committee should know, October We obtain governance and incentives data from Equilar, which is similar to the ExecuComp database in that it provides executive compensation and equity holdings data collected from annual proxy filings (Form DEF 14A) with the SEC. We use Equilar data because it provides more than twice as many annual observations than does ExecuComp

12 Table 1 Panel A provides descriptive statistics for all of the variables used in our analysis including measures of tax avoidance, governance, incentives, and the control variables. Table 1 Panel B compares characteristics of our sample to those of the Compustat universe for fiscal year 2009, which is the year that has the largest representation in our sample. Panel B shows that the firms in our sample are represented in each of the Barth et al. (1998) industry groups and, on average, are larger and more profitable than those in the Compustat universe. 4. Research Design Quantile regression provides the opportunity to draw richer inferences beyond those that can be drawn from traditional OLS regressions, which only provide inferences about the association between X and the conditional mean of Y. Quantile regression, instead, provides evidence about the association between X and any specified conditional percentile of the Y distribution. This allows researchers, for example, to understand how X is associated with the Y distribution at its right and left tails which is arguably where the relevant research focus might be. Hao and Naiman (2007), note, for example that the focus on the central location has long distracted researchers from using appropriate and relevant techniques to address research questions regarding noncentral locations on the response distribution. Using conditional-mean models (e.g., OLS regression) to address these questions may be inefficient or even miss the point of the research altogether. A set of equally spaced conditional quantiles can characterize the shape of the conditional distribution in addition to its central location. This point may perhaps be best illustrated through an example. Assume a government researcher is interested in learning whether education affects pay to help inform public education policy. When estimating OLS regression at the conditional mean, the researcher observes a small positive association between

13 education and pay. The researcher might infer that the average effect of education on pay is modest. The policy inferences from OLS, therefore, might be that further education investment generates only modest pay benefits. If the researcher, however, estimates the regression at the conditional 5 th percentile of the pay distribution, he or she might learn that the association between education and pay is much larger. This might suggest that the effect of education on the distribution of pay is much larger in the poorest communities, where the focus of education policy might be more salient. For our setting, quantile regression allows us to consider whether the association between tax avoidance and certain governance characteristics varies based on where the association is examined along the tax avoidance distribution. As discussed in Section 2, we hypothesize that knowledgeable and independent boards actively engage managers regarding tax choices when the level of tax avoidance is extremely high or low. Thus, any relation between managerial incentives or corporate governance and tax avoidance should be stronger in the tails of the tax avoidance distribution. From a methodological perspective, this prediction implies that the impact of managerial incentives or specific governance mechanisms may alter the shape of the entire tax avoidance distribution, rather than simply shift only the mean or median of its location. Consistent with some of the results from prior studies, it is quite possible that the impact of governance is negligible at the mean or median, but very strong at other points of the tax avoidance distribution. Since traditional regression estimation methods (e.g., OLS) are not equipped to detect such a distributional shift, we use quantile regression as our primary method of estimation. The basic intuition for quantile regression in the context of our hypotheses is illustrated by the diagram in Figure 1. For example, the treatment effect of a particular governance mechanism

14 may rotate the conditional tax avoidance cumulative distribution function (CDF) around the median in a counterclockwise manner. Such an effect would manifest as a negative (positive) coefficient on governance in the upper (lower) tail of the tax avoidance distribution. However, the coefficient on governance at the median (i.e., the central location) would be zero. Since traditional OLS or median regression only estimates the treatment effect at the center of the distribution (i.e., the mean and median, respectively), these techniques cannot detect such shifts in the distribution of interest. In contrast, quantile regression is much more general and is designed to characterize changes in both the location and shape of the distribution of interest. To better understand the mechanics behind quantile regression, recall that OLS regression solves the problem: min ( ), Quantile regression has similar characteristics which can be seen by the special case of regression at the 50 th percentile (i.e., median regression): min, To estimate the association at any other specified percentile of the Y distribution, quantile regression relies on an asymmetric weighting problem that relies on a parameter c, defined by a check function: min R, R ( ), where is the quantile (i.e., percentile) of interest, u is the error estimate, ( )=( 1 0 +(1 )1 <0 ) =( 1 <0 ),

15 and 1[*] is the indicator function (i.e., a dichotomous variable that equals one if the bracketed condition regarding u is true and equals zero otherwise) Quantile regression specification To test our research hypotheses, we estimate the following specification using quantile regression. 11 TaxPosition i,t = β 0i,t + β 1 LogNumFinExp i,t-1 + β 2 PctIndep i,t-1 + β 3 LogNumDirs i,t-1 + β 4 LogCEOPortDelta i,t-1 + β 5 LogCEOPortVega i,t-1 + β 6 CFOps i,t + β 7 LogMVE i,t + β 8 LogForAssets i,t +β 9 GeoComp i,t + ε i,t, (1) TaxPosition is one of two proxies that measure a firm s level of tax avoidance for a given year. Our first proxy for tax avoidance is EndFin48Bal, which we measure as the firm s ending balance of its uncertain tax benefit account (Compustat item TXTUBEND), scaled by total assets. 12 Our second tax avoidance proxy is TAETR, which we calculate as the difference between the firm s three-year average (i.e., mean) GAAP effective tax rate (hereafter, ETR, 10 Imbens, Guido W. and Jeffrey M. Wooldridge, Lecture Notes 14, Summer 07, available at 11 Following Koenker and Hallock (2011), we both report and plot graphs of the coefficient estimates at decile intervals of the tax avoidance distribution to show how the relationship changes across the support of the distribution. We also report OLS coefficient estimates of the conditional mean for comparative purposes. 12 There is considerable debate in the tax literature about the appropriate measure of tax avoidance. De Waegenaere et al. (2010) show that that best proxy for tax avoidance activity found in the financial statements is often the FIN48 reserve. However, we assess the sensitivity of our results to a number of alternative tax avoidance measures, including a modified version of Frank et al. s (2009) DTAX measure of discretionary permanent tax differences and EffUTB, which is the portion of the firm s uncertain tax benefit that, if reversed, would increase the firm s effective tax rate. Frank et al. (2009) compute DTAX as the residual from a regression of permanent tax differences on measures of intangible assets, income of unconsolidated subsidiaries, minority interest, state tax burdens, changes in NOLs, and lagged permanent differences. We modify this computation by including Oler et al. s (2007) measure of foreign assets (LogForAssets) to control for the existence of multinational operations. By including foreign assets in the first stage, we attempt to control for ETR differentials that result from ordinary overseas operations. Without this modification, DTAX would suggest that firms with extensive foreign operations or foreign operations in low tax jurisdictions are more aggressive tax planners. All reported results are robust to excluding foreign assets and measuring DTAX exactly as described by Frank et al. (2009). Our DTAX measure is scaled by the firm s average total assets during the period. These alternative tax measures produce results (untabulated) that are very similar to those from our primary measures of tax avoidance. Thus, our tabulated results are not unique to any specific measure of tax avoidance

16 computed as the firm s total tax expense scaled by pre-tax income) and the three-year average GAAP ETR of the firm's size and industry peers (i.e., those in the same quintile of total assets in the same Fama-French 48 industry groups). This measure of tax avoidance captures crosssectional variation in firms total tax planning (including both timing and permanent differences), and benchmarks a given firm s tax aggressiveness relative to that of similar-sized firms in the same industry (see Balakrishnan et al., 2012). LogNumFinExp is the natural logarithm of one plus the number of financial experts on the board of directors in the previous year (as indicated by RiskMetrics). We use the number rather than the proportion of financial experts on the board because we believe it better measures the potential for incrementally greater tax expertise. 13 PctIndep is the percentage of independent directors to total directors sitting on the board in the previous year (as indicated by Equilar). LogNumDirs is the natural logarithm of one plus the number of total directors sitting on the board in the previous year (as indicated by Equilar). LogPortDelta is the natural logarithm of the (risk-neutral) dollar change in CEO equity portfolio value for a 1% increase in stock price (Core and Guay, 2002) during the previous year; LogPortVega, is the natural logarithm of one plus the (risk-neutral) dollar change in CEO equity portfolio value for a 0.01 increase in annual stock return volatility (Core and Guay, 2002) during the previous year. 14 We control for fundamental economic determinants of firms tax positions by including variables commonly found in the tax literature that affect firms costs, benefits, and opportunities to engage in tax avoidance (e.g., 13 For example, we believe there is likely to be greater tax expertise on a board with two financial experts out of ten total board members than there would be on a board with one financial expert out of five total board members. 14 The value of a CEO s stock and restricted stock is assumed to change dollar-for-dollar with changes in the price of the underlying stock. The value of a CEO s stock options is assumed to change according to the option s delta (vega), which is the derivative of its Black-Scholes value with respect to the price (volatility). Annualized volatility is calculated using continuously compounded monthly returns over the previous 36 months, with a minimum of 12 months of returns. The risk-free rate is calculated using interpolated interest rate on a Treasury note with the same maturity, to the closest month, as the remaining life of the option multiplied by 0.70 to account for the prevalence of early exercise. Dividend yield is calculated as the dividends paid over the past 12 months scaled by the stock price at the beginning of the month

17 cash flow from operations, the amount of foreign assets, the market value of equity, and the degree of geographic complexity). CFOps is computed as cash flow from operations divided by average total assets; LogMVE is the natural logarithm of market value of equity; LogForAssets is the natural logarithm of total foreign assets; GeoComp is a revenue-based Hirfindahl-Hirschman index that captures within-firm geographic segment complexity (Bushman et al., 2004); and i and t index firm and fiscal year, respectively. Panel C of Table 1 reports the mean level of CEO equity incentives (i.e., CEOPortDelta and CEOPortVega) for each decile of the tax avoidance distribution (where tax avoidance is measured as EndFin48Bal). Evidence in Table 1 Panel C indicates that CEO equity incentives tend to increase with the level of tax avoidance. Finally, Table 1 Panel D reports Pearson correlations among the variables. This correlation matrix shows that tax avoidance is associated with cash flow from operations, the amount of foreign assets, and the degree of geographic complexity, which underscores the importance of controlling for these factors in our multivariate specifications. 5. Results Table 2 and Figure 2 present our primary evidence on the relation between firms governance including CEO equity incentives and tax avoidance. We first consider the degree to which CEO incentives (LogCEOPortDelta and LogCEOPortVega) affect the level of tax avoidance. OLS estimates of the conditional mean of tax avoidance presented in Panel A of Table 2 do not provide evidence of a statistical association between TaxPosition and LogCEOPortDelta. Consistent with Rego and Wilson (2012), OLS estimates provide evidence of

18 a positive association between TaxPosition and LogCEOPortVega, particularly when tax avoidance is measured with EndFin48Bal. Figure 2 Panel A plots quantile regression coefficient estimates from Table 2 Panel A and provides a more comprehensive depiction of the relation between our measures of CEO equity incentives and TaxPosition. In general, the patterns depicted in Figure 2 Panel A indicate that the relation between CEOs equity incentives and tax aggressiveness is mostly positive and increasing in magnitude in the right tail of the distribution (particularly with respect to LogCEOPortVega), which is consistent with unresolved agency problems with respect to tax avoidance. Figure 2 Panel B and Table 2 Panel B both provide evidence about the association between tax avoidance and board expertise and independence. We focus first on LogNumFinExp which measures the degree of financial sophistication on the board. We expect that more sophisticated boards will better understand when a firm might be over- or under-investing in tax avoidance, and therefore will have a greater effect on the firm s tax avoidance at extreme levels of tax avoidance. We therefore expect to observe a positive relation between LogNumFinExp and TaxPosition in the left tail a negative relation in the right tail of the tax avoidance distribution. OLS estimates in Panel B of Table 2 do not provide evidence of a statistically significant relationship between TaxPosition and LogNumFinExp for either tax avoidance measure. In contrast, for both tax avoidance proxies, Figure 2 shows that the relation between the financial expertise of the board and tax avoidance varies at different points of the tax avoidance distribution. Specifically, the association between boards financial sophistication and tax avoidance is positive at low levels of the tax avoidance distribution but negative at high levels of tax avoidance. This result is consistent with our hypothesis and suggests that boards financial

19 sophistication does not have a uniform (i.e., constant) effect on tax avoidance. Moreover, the quantile regression estimates indicate that the OLS estimates are not representative of any effect at other points of the distribution. Next we consider PctIndep, which measures board independence, and has been used as a measure of the degree of board monitoring and oversight of management. Desai and Dharmapala (2006) suggest that agency problems may arise with respect to tax avoidance which could occur, for example, if shareholders net benefits from tax avoidance are concave and the net benefit to managers is linear. If more independent boards can better identify and, in turn, mitigate these agency problems, then we expect to observe a positive (negative) relation in the left (right) tail of the tax avoidance distribution. OLS estimates presented in Panel B of Table 2 provide no evidence of aan association between TaxPosition and PctIndep. However, similar to our results for LogNumFinExp, the quantile regression coefficient estimates in Table 2 and the accompanying graph of these estimates in Figure 2 suggest that the relation between board independence and tax avoidance varies at different points of the tax avoidance distribution. This pattern is consistent with our hypothesis that more independent boards ameliorate over- and under-investment in tax avoidance, which may be symptomatic of unresolved agency problems with respect to tax planning. We acknowledge that firms observed governance structures are not randomly assigned (i.e., are not exogenous). Therefore, we cannot definitively rule out potentially confounding inferences like reverse causality. Under the assumption of reverse causality, it is possible that tax positions are determined and then governance structure is developed to support these tax positions. For example, firms may rely on more financially sophisticated boards to pursue the

20 most aggressive tax positions. If this is the case, we would expect to observe a positive association between LogNumFinExp and TaxPosition in both tails, with a larger positive association in the right tail of the tax avoidance distribution. Our results do not support this alternative inference, so we do not believe that reverse causality is confounding our inferences. Likewise, we cannot definitively rule out the possibility that the observed relationship between governance characteristics and tax avoidance may be affected by omitted factors that are correlated with the (non-random) design of firms governance mechanisms. We control for common determinants of firms governance structures, so we believe that it is unlikely that our primary results are attributable to factors that are correlated with both the design of firms governance mechanisms and their tax avoidance (i.e., correlated omitted variables). We also believe it is unlikely that there exists some omitted correlated variable that would induce both positive associations at the lower and negative associations at the upper tails of the tax avoidance distribution, but we cannot rule this possibility out entirely. 6. Supplemental Analysis 6.1. Sensitivity Analysis: Alternative Investment Measures As discussed earlier, tax avoidance can be considered one of several investment options available to the firm. One way to strengthen the confidence in our inferences is to assess whether we observe similar association patterns between investment and governance characteristics across the investment distribution for other investment choices, such as capital expenditures (CAPEX) or R&D. One might expect a similar potential for under- or overinvestment with CAPEX or R&D. If so, the association between investment and board sophistication may vary predictably as a function of where the association is measured on the

21 investment distribution. Consistent with this hypothesis, we observe quantile regression coefficient patterns (untabulated) similar to those reported in Figure 2, when Tax Avoidance is replaced with CAPEX or R&D as the dependent variable. 15 This suggests that these governance characteristics may also ameliorate agency issues associated with other investment options. 6.2.Sensitivity Analysis: Changes in Tax Avoidance If sophisticated boards recognize that their firms are under- or over-investing in tax avoidance, one might expect to observe tax avoidance corrections in subsequent periods. To examine this question, we partition each observation into ENDFIN48BAL t-1 quintiles. We then examine whether ENDFIN48BAL t exhibits predictable corrections as a function of the number of financial experts on the board or the percentage of independent directors on the board. For the lowest quintile of ENDFIN48BAL t-1, we observe a change (untabulated) of in the next year when there are at least three financial experts on the board. We observe a smaller (albeit not statistically different) change of when there are fewer financial experts on the board. Likewise, for the highest quintile of ENDFIN48BAL t-1, we observe a change of in the next year when there are at least three financial experts on the board. We observe a statistically significant smaller change of when there are fewer financial experts on the board. For the lowest quintile of ENDFIN48BAL t-1, we also observe a change of (0.0009) in the next year when the percentage of independent directors on the board is greater than (less than) the sample median. These changes are not statistically different from each other. For the highest quintile, we observe a ( ) change in the next year when the percentage of independent directors on the board is greater than (less than) the sample median. These changes are weakly significantly different. Collectively, these results provide some evidence of 15 We do not observe similar patterns when ROA is the dependent variable, which suggests that we are not simply documenting quantile regression coefficient patterns that are spurious or mechanically determined by research design choice

22 predictable tax avoidance corrections associated with more sophisticated or independent boards Desai and Dharmapala (2006) Desai and Dharmapala (2006) present evidence that their measures of governance and executive incentives have an interactive effect on the average level of firms tax aggressiveness. Specifically, they provide evidence (in their Table 4) that there is no statistical relation between their measure of tax aggressiveness and the ratio of annual stock option grant value to total compensation for the firm s top five executives in well governed firms (i.e., those that have a low governance index score). 16 However, they report a significant negative association between their measure of tax aggressiveness and the ratio of stock option grant value to total compensation for the firm s top five executives in poorly governed firms (i.e., those that have a high governance index score). We can generally replicate this result (i.e., Table 4 Column (4) of Desai and Dharmapala, 2006) in our sample. 17 Desai and Dharmapala (2006) discuss why it is not clear whether equity compensation provides incentives for managers to engage in more or less aggressive tax strategies. They also discuss how it is reasonable to expect that the degree to which equity-based compensation provides incentives for tax aggressiveness is a function of the firm s other governance mechanisms since they can serve as complements and substitutes for each other. However, they do not attempt to either model this complexity in their research design or consider that effects may vary across the tax avoidance distribution. Therefore, we re-estimate Desai and 16 Desai and Dharmapala (2006) measure tax aggressiveness as the residual from a regression of the difference between book and tax income on total accruals (see their discussion, p ). They also rely on the Gompers et al. (2003) G-index, which is primarily a measure of shareholder rights, as their measure of governance quality. 17 Tabulated results in Desai and Dharmapala s Table 4 Column (4) do not include the interaction between equity incentives and governance. Rather, they estimate the analysis separately for well-governed and poorly-governed firms. Hence, we are unable to formally compare our finding of a statistically significant coefficient on the interaction with the results from their specification

23 Dharmapala s (2006) specification (i.e., their Table 4 Column (4)) using quantile regression to determine whether the conditional mean effect that they document is representative of the effect at other points in the distribution or whether the relationship instead varies as suggested by our results presented in Figure 2 and Table 2. Figure 3 and Table 3 provide OLS and quantile regression estimates of Desai and Dharmapala s specification for our sample. 18 Using OLS estimation, we are generally able to replicate Desai and Dharmapala s result of a negative (marginally significant) coefficient for the ratio of stock option grant values and a positive (but insignificant) coefficient when this ratio is interacted with an indicator for whether the firm is well governed. Quantile regression estimates, however, show that the interactive effect of executive equity grants and well governed firms varies across the tax avoidance distribution. Specifically, we find that this relation is negative in the right tail of tax the tax avoidance distribution, which suggests that both equity compensation and corporate governance may, in fact, mitigate aggressive tax positions. These results are consistent with our results in Figure 2 and Table 2, but are at odds with Desai and Dharmapala s (2006) conclusion. 7. Conclusion We examine the relation between firms corporate governance mechanisms (including managers equity incentives) and their level of tax avoidance. We are interested not only in the average level (i.e., conditional mean) of this relation, but also how it varies across the tax avoidance distribution. We are especially interested in examining effects in the tails of this distribution, which capture relatively extreme levels of tax avoidance. In contrast to prior studies 18 For completeness, we include a main-effect well-governed firm dichotomous variable in this estimation, which was inappropriately omitted in Desai and Dharmapala (2006)

24 that only examine this relation at the conditional mean, we estimate quantile regression, which allows us to determine whether (and how) the relation varies across the tax avoidance distribution. The ability to observe shifts in the relation is particularly important in this setting because firms governance mechanisms are likely to exhibit different relations with tax avoidance at different points in the distribution if the net benefits of tax avoidance exhibit diminishing marginal returns. Consistent with the hypothesis that management expects greater personal payoffs from increased tax avoidance (e.g., Rego and Wilson, 2012), we find evidence that risk-taking equity incentives are positively related to tax avoidance and that this association is greater in the right tail of the tax avoidance distribution. This result is consistent with our conjecture that relatively high levels of risk-taking equity incentives have the potential to motivate managers to overinvest in tax avoidance relative to the level desired by shareholders. We also find evidence that board financial sophistication and independence exhibit a positive (negative) relationship with tax avoidance in the lower (upper) tail of the tax avoidance distribution. This finding is consistent with the hypothesis that more sophisticated and independent boards recognize that there are concave net benefits to tax aggressiveness and mitigate potential agency problems. Collectively, our study provides evidence that yields a richer understanding of how managerial incentives and certain corporate governance mechanisms affect firms tax avoidance. Perhaps more importantly, our study provides evidence that is consistent with the existence of concave net benefits of tax avoidance. Although our results do not speak to specific costs that are associated with aggressive tax avoidance, our empirical methodology and evidence should help researchers better understand what prevents all firms from reaching the corner solution of maximum tax avoidance (i.e., a zero effective tax rate). Empirical methods that estimate only the