theoretical and empirical
examination
Jos Offerein (s2222647)
University of GroningenResearch Master Thesis
August 14, 2017
Abstract
In this study, corporate short-termism is connected to corporate tax avoidance through a theoretical model that draws on Desai and Dharmapala’s (2006) model on tax avoidance and managerial diversion. Among other things, the model predicts a positive relation between short-termism and corporate tax avoidance, conditional on a restriction. The insights from this model are subsequently tested empirically, using measures from short-termism that are derived from textual analysis of conference calls. Furthermore, a new, composite measure of tax avoidance is employed to study the proposed relations. Although the evidence is not robust, there is a positive relation between short-termism and corporate tax avoidance in some specifications. Concluding, further research is thus needed to assess if and how the model suggested in this study can generate further insights into the potential for principals to request socially harmful actions from managers.
I.
Introduction
T
ax avoidance1by firms currently poses an important challenge for developed economies. The perceived unfairness of large corporations paying virtually no taxes has rattled many citizen action groups and politicians on a national and international level. Recent research shows that executives within firms indeed actively consider tax consequences when making decisions (Graham, 2003; Alvarez and Marsal, 2012; Graham et al., 2014). These considerations are by no means trivial. As Slemrod (2004) notes, the US collected 6.4 percent of GDP in corporate tax revenues in the year 1951. In the early years of the 2000s, this was reduced to 1.5 percent. Since then, there has been some volatility in tax revenues, but on average still only approximately 2 percent of GDP is collected in taxes from corporations in the U.S.Although exact numbers on the absolute amount of tax avoidance by firms are hard to come by, essentially because firms do not report how much they avoid, there are several estimates available. Slemrod (2004) reports in his table 1 the estimated loss due to misreporting. This estimate steadily increases from $60.9 billion in 1988 to $146.8 billion in 2000. Recent estimates published by the Tax
Justice Network2(TJN) show that this trend has persisted: in the past few years, $188.8 billion is approximately lost due to corporate tax avoidance in the U.S. each year. Global losses amount to $500 billion according to the information of the TJN, whilst the IMF estimates the global loss in revenues to be even larger: $600 billion. In sum, these practices are thus costly for societies. In addition, research on corporate tax avoidance has only emerged relatively recently, which results in an incomplete understanding of its determinants and role in the corporate decision-making process.
Past research has empirically examined several potential contributors to this large-scale dodg-ing of taxes3. Most of the theoretical research to date has adopted the agency perspective on corporate tax avoidance, which was introduced in Chen and Chu (2005) and Crocker and Slemrod (2005). Even more influential was the theory developed by Desai and Dharmapala (2006), who emphasized the role of equity incentives for executives and the broader governance environment as a determinant of tax avoidance. Although the approach in this paper incorporates and re-examines insights from the literature that subsequently arose after these seminal papers, I also attempt to extend the literature by introducing time preferences of shareholders. In order to do so, catering theory (Stein, 1989; 1996) is invoked as an explanation of how time preferences might affect tax avoidance decisions for firms. Catering theory is, briefly stated, the adherence of managers to the time orientation of their average shareholder. Thus, firms with a shareholder base that plans on being invested for a prolonged period of time in this particular firm, induces managers to take decisions oriented towards long-term value creation. On the other hand, firms with a short-term oriented shareholder base prefer a higher value of the firm in the short-term, with little regard for the long-term financial well-being of the firm.
This type of short-termism is currently also under scrutiny of researchers and policymakers4, with recent evidence providing first indications that this short-term orientation is indeed causing a downward pressure on long-term value creation (Souder et al., 2016; Flammer and Bansal, 2017). The aim of this paper is to connect these two harmful phenomena, tax avoidance and short-termism, much like previous research has done for managerial diversion and corporate tax avoidance (Chen and Chu, 2005; Desai and Dharmapala, 2006). The main challenge in doing so is to develop theoretical underpinnings of such interrelations. By developing a simple model based on Polk and Sapienza (2009) and Desai and Dharmapala (2006), such a theory is derived.
Solving the model and obtaining comparative static effects for the parameters short-termism and equity related compensation eventually yields four testable propositions, for which the direction of the comparative static effects hinges on two conditions. First, a proposition on the main contribution of this paper, the relation between short-termism and corporate tax avoidance, is obtained. From the mechanics of the model and a derived condition, a positive relation is predicted. Second, since the model includes the degree of managerial diversion (Desai and Dharmapala, 2006), the effect of short-termism on managerial diversion is also obtained. For this relation, a positive effect is also derived, given that the same condition holds. In sum, the subsequent empirical tests will either show that the above propositions both hold, and therefore that condition 1 is satisfied, or, that neither will hold in the context of this model.
The second set of propositions relate to the effect of equity incentives on tax avoidance and managerial diversion. In line with the recent empirical work on equity incentives and tax avoidance (Philips, 2003; Minnick and Noga, 2010; Rego and Wilson, 2012; Armstrong et al., 2015), the model
2See http://www.taxjustice.net/2017/03/22/new-estimates-tax-avoidance-multinationals/ for the source of the num-bers.
3For reviews of the topics in corporate tax avoidance that were empirically examined, see Hanlon and Heitzman (2010) or Lietz (2013).
allows for a positive effect of equity incentives on tax avoidance whilst equity incentives are also predicted to decrease managerial diversion. This result closely corresponds to the first (but rejected) proposition in Desai and Dharmapala (2006). Note that these propositions do not attempt to add new insights to the literature, they merely serve to complete the model and verify whether these relations hold empirically in the dataset used to test the first two propositions.
Besides the theoretical model that is developed here, this study also provides further insights into the potentially harmful role that principals might play in the corporate landscape. As Flammer and Bansal (2017) and Souder et al. (2016) documented, short-termism is detrimental to firm value. Therefore, social welfare is reduced as a consequence of the short-term orientation of shareholders. As Harford et al. (2015) show, a long time horizon of the shareholder base is also beneficial for firm decision-making and governance quality, further limiting the scope for socially undesirable actions by agents. In sum, this study potentially adds tax avoidance to the list of undesirable actions induced by principals rather than agents, which strengthens the case for countermeasures to limit shareholder short-termism, and more generally, other detrimental behaviour by principals. Apart from the theoretical contributions, further additions to the literature lie in the subsequent tests. Specifically, the empirical procedure contains at least two further contributions. First, the measurement procedure for corporate tax avoidance is a novel one, combining five5 different measures that were previously used in tax avoidance research, but each measure different aspects of tax avoidance, into one composite index. Specifically, measures of effective tax rates, use of tax havens, the balance of uncertain tax benefits, predicted participation in tax sheltering, and temporary book-tax differences are all combined in a single measure by calculating z-scores and averaging these out.
Second, the recently popularized method of textual analysis is employed to measure short-termism. Following the pioneering work of Brochet, Loumioti and Serafeim (2015) and DesJardine and Bansal (2015), these measures, which are based on automatically counting words that are related to either the short or the long-term, are used to approximate the short-termism that the shareholder base introduces into the firm. To the best of my knowledge, this study is the first to introduce textual analysis to the field of corporate tax avoidance.
The results of the empirical tests are mixed. On the one hand, there is some evidence that short-termism, as measured by the Brochet et al. (2015) construct, positively affects corporate tax avoidance. On the other hand, robustness tests using real activities management or the alternative measure of DesJardine and Bansal (2015), show no significant effects, although throughout the tests the desired sign is consistently found. The relation between short-termism and managerial diversion has the predicted effect, but only for the real activities variable: higher short-termism increases managerial diversion. Moreover, in robustness tests with alternative measurement of short-termism, this relation becomes negative, which is a puzzling result. The results on equity incentives are also inconclusive, with the main results producing the expected positive relation between equity incentives and short-termism, but any relation appears to be absent for managerial diversion, and in robustness tests with alternative measurement of short-termism neither relation is found.
Finally, alternative econometric techniques yield no results when an attempt is made to counter endogeneity. On the other hand, a quantile regression describes an interesting pattern for the effect of short-termism on tax avoidance across the tax avoidance distribution. Apparently, the more taxes a firm avoids, the stronger the effect of short-termism becomes. In sum, there is some evidence that short-termism affects corporate tax avoidance. The model that is proposed
in this paper however seems to be inadequate in capturing the entire phenomenon, since the crucial conditions that determine the sign of the comparative static effects are occasionally violated. Therefore, further research is necessary to obtain a complete theory of the potential relation between the two socially undesirable practices of tax avoidance and short-termism.
The paper has the following outline: the second section discusses related literature, on both the short-termism and tax avoidance dimension. Next, the third section introduces the model and discusses its empirical implications. Subsequently, the data and methodology for the empirical tests are discussed, before the fifth section presents results of the empirical analysis. Finally, the sixth section discusses the results and theoretical implications, and concludes the paper.
II.
Literature review
This paper is related to two topics in the literature, the literature on agency problems and tax avoidance, and the literature that studies short-termism, particularly its consequences. Since there is no previously established explicit link between these two strands of the literature, they will be discussed in turn, followed by a brief discussion of the literature on shareholder characteristics and tax avoidance. The topics discussed in this section then serves as a foundation for the model in the next section.
Corporate tax avoidance in agency situations
Although the problem of non-compliance with tax obligations is as old as taxes themselves (An-dreoni, Erard and Feinstein, 1998), only relatively recently has tax avoidance attracted attention from researchers. The seminal paper by Allingham and Sandmo (1972) on tax evasion by indi-viduals demarcates the start of this literature. However, the avoidance decision of corporations is arguably different from the evasion decision by individuals, due to the separation of ownership and control and the potential problems this causes (Jensen and Meckling, 1976). Slemrod (2004) acknowledges this, and proposes to study corporate tax avoidance in a principal-agent framework. Subsequent literature followed this recommendation, studying different aspects of this relationship, such as: devising an optimal contract to incentivize managers to engage in tax avoidance (Chu and Chen, 2005), whether the principal or the agent should be penalized for evasion (Crocker and Slemrod, 2005), the role of specialists in tax avoidance (Lipatov, 2012) and who should be liable for tax avoidance (Biswas, Marchese and Privileggi, 2013). However, theoretical work has to date not yet explored potential consequences of particular characteristics the shareholder base might have, whilst shareholders have been shown to have substantial influence on firm decision-making (see Boyd and Solarino (2016) for a review). This paper aims to contribute to this strand of the literature.
For the purposes of this study, the empirical evidence that emerged as a response to the contro-versial theory of Desai and Dharmapala (2006) is interesting, since it refutes the complementarity of tax avoidance and managerial diversion6. Desai and Dharmapala (2006) developed a model in which an increase in equity incentives for management would decrease managerial diversion and increase tax avoidance, or it would decrease both. The subsequent empirical analysis revealed that the latter proposition holds. An increase in equity incentives decreases both managerial diversion and tax avoidance, but only for poorly governed firms, where equity incentives could still serve to decrease managerial diversion. Several authors subsequently examined whether this result indeed holds, using more specific measures of equity incentives, governance quality and tax
avoidance (Minnick and Noga, 2010; Armstrong, Blouin and Larcker, 2012; Rego and Wilson, 2012; Armstrong et al. 2015). Almost unanimously across the different specifications and techniques employed, these authors conclude that a higher sensitivity to the firm’s stock price7increases corporate tax avoidance. This finding is thus fairly robust in the empirical literature, and therefore the model developed below incorporates this finding into a model that is otherwise relatively similar to that of Desai and Dharmapala (2006).
A related topic in the literature on tax avoidance to which this paper might contribute is what has come to be known as the undersheltering puzzle (Weisbach, 2002). Basically, it has been argued that given the availability, relative ease and low costs of tax shelters, there appears to be no good reason why firms do not employ these tax shelters to a greater extent. Based on the model derived in the next section, shareholder time orientation might substantially affect the perceived costs of tax avoidance, hence pushing down the desired level of tax avoidance a set of given shareholders feels comfortable with, thus giving rise to undersheltering. However, Gallemore, Maydew and Thornock (2014) show that there appear to be no significant reputational costs associated with tax avoidance, and hence this apparently cannot be an explanation for the undersheltering puzzle, nor could it lead to shareholders expecting costs from tax avoidance.
However, Hanlon and Slemrod (2009), and Graham et al. (2014) find evidence that reputational costs are in fact real (by declining stock prices) and an important considerations for CFOs in the tax avoidance decision. Therefore, if the market operates based on expectations, whether reputational or other costs exist might not matter, the fact that executives and shareholders expect them to exists validates taking them into account. Moreover, recent developments in the political and legislative landscape could lead shareholders to expect that possibilities for tax avoidance structures will reduce or diminish in its current form. Therefore, investing now in constructions that will have to be dismantled in the upcoming years seems to be an unprofitable strategy for anyone who is invested for the long-term. Short-term investors (given surmountable upfront investment costs) might however still be interested, since they will divest long before such threats to tax avoidance structures will materialize.
Short-termism and catering theory
This paper relies quite firmly on catering theory to hold. This theory was first developed (implicitly) by Stein (1996), although some foundations were already laid in Stein (1989). Catering theory unfolds in the interplay between shareholders and managers. More precisely, managers must take investor perceptions into account when they are optimizing short-run firm value. This implies that a short-term oriented investor base induces vastly different behaviour from managers than a long-term oriented investor base. With a long-term investor base, the manager can follow his own valuations and strategies because over the long-term any faulty investor perceptions of the firm will resolve. If the investor base is short-term oriented, the manager will take a different approach to optimizing the firm value, taking into account the misperceptions of the shareholders. Recently, this was shown to result in a lower firm value in the long-run (Souder et al. 2016; Flammer and Bansal, 2017).
Substantial empirical evidence exists in favour of catering theory, although there are also results that point to opposite conclusions. One of the implications that was tested repeatedly is the prediction that market valuations should matter for firm investment policies, otherwise, catering theory could be rejected. Early tests of this proposition yielded mixed results (Barro, 1990; Morck, Shleifer and Vishny, 1990; Blanchard, Rhee and Summers, 1993; Chirinko and Schaller, 2001). More recent attempts to examine whether catering theory is borne out by the data are more
consistent, all providing evidence in favour of (some form of) catering theory (Polk and Sapienza, 2009; Derrien, Kesckés and Thesmar, 2013, Brochet et al., 2015). Hence, there is evidence that the time orientation of firms is aligned with the time orientation of shareholders.
Previous literature has also attempted to develop insights in the consequences of short-termism. The most important result of these efforts is probably the finding that short-termism is actually a cause for lower overall firm value, as Flammer and Bansal (2017) show using a regression discontinuity design. Similar evidence is obtained in Harford, Kesckés and Mansi (2015), for a range of outcomes which ultimately also result in a higher firm value. These authors examine whether long-term investors improve corporate decision making and they conclude that managerial misbehaviour is decreased and governance is improved. Furthermore, a range of financing and investment decisions is affected. Brochet et al. (2015) use, similar to this study, conference calls to measure the short-termism of firms. They find that short-termism increases earnings and real activities management. Similar results are also obtained by Derrien et al. (2013) and Harford et al. (2015), although with a substantially different approach to measuring short-termism. Derrien et al. (2013) establish that longer investor horizons decrease mispricing, implying that short-termism acts to inflate short-term profits unrealistically, therefore causing short-term oriented firms to become overvalued. In sum, there is substantial evidence that short-termism influences firm decision-making.
As discussed, firm management takes tax considerations into account when making decisions (Graham, 2003; Alvarez and Marsal, 2012; Graham et al., 2014). Conversely, as just discussed, managers also take time preferences of their shareholders into account, by translating them into real effects on investments and other decisions (Brochet et al. 2015; Harford et al. 2015). Therefore, when determining the optimal policy with respect to taxation, one would expect that the time orientation of the firm influences this, conditional on the existence of an inter-temporal trade-off, that is, tax avoidance is more profitable on either the short-term or the long-term. To develop intuition on which of these is most plausible, the literature on tax avoidance and other shareholder characteristics is briefly reviewed below before turning to the model.
Ownership and tax avoidance
To date, some research has been done to determine the effects of other ownership characteristics on corporate tax avoidance. Chen et al. (2010) examined whether family firms avoid more taxes compared to non-family firms. They conclude that this is not the case, if a family holds a large part of the shares, firms exhibit lower tax avoidance. Although there might be a number of reasons why this prevails, a possible explanation is that family firms are by definition dominated by a long-term shareholder. Following the reasoning outlined above, such a shareholder would favour lower tax avoidance, and consequently the manager will set a lower degree of tax avoidance, which is in line with the findings of Chen et al. (2010). Further circumstantial evidence in the tax avoidance literature on shareholder characteristics is provided by Cheng et al. (2012). These authors investigate the effect of hedge fund intervention on corporate tax avoidance from the underlying premise that hedge funds often become large shareholders upon intervention and can therefore push for increases in tax avoidance. The data indeed confirm that after a hedge fund intervention, firms exhibit greater tax avoidance. Since hedge funds are usually considered to be short-term investors, this could imply a preference for greater tax avoidance if the shareholder base is short-term oriented.
argue, privately held firms do not face the scrutiny of public firms, and moreover they possibly face lower costs upon detection. This could justify a higher level of tax avoidance for the subset of private firms, even if their shareholder base is long-term oriented8.
III.
A simple model
The model
Comparable to Desai and Dharmapala (2006), firms are assumed to generate income exogenously. That is, managers are assumed to have limited influence on the actual performance of the firm, at least not significantly more than other managers. The only differences that will affect final firm value are the manager’s willingness to avoid taxes and tendency to divert income. To model this, suppose that all firms generate some income Ytin each period t. Let Ytbe a random variable with
probability density function f(Yt)9. Note that Ytis the before-tax income; after it is realized there
are only two decisions to be made: how much to report to the tax authorities10 and whether the manager will divert part of the realized income for his own gain.
To model these two decisions, denote the avoidance decision by α, which implies that α is the proportion of income that is reported to the tax authorities, so necessarily, α∈ [0, 1]. Similarly, β reflects the proportion of income that is not diverted, with(1−β)the proportion the manager
will divert, and consequently β∈ [0, 1]. Finally, a tax rate τ is required, to reflect the proportion of reported income that has to be transferred to the tax authorities. In sum, the expected value of the firm at time t is then:
Vt= (1−ατ)β ¯Yt, (1)
with ¯Yt=EYt=R Ytf(Yt)dYt. Note that it is assumed that α, β and τ are constant over time, that
is, managers and the tax authorities choose values for these parameters and do not change them over time.
However, this specification of firm value Vtdoes not account for any costs that might result from
tax avoidance. There can be a range of possible costs and benefits from tax avoidance, including: reduced taxes paid, lower compliance costs, higher compliance costs (due to a complex corporate structure), costs for a tax avoidance structure, reputational costs upon discovery, potential fines from the tax authorities, potential voluntary additional payments to the tax authorities in order to restore the company’s image, etcetera. The key point of this model is to illustrate how time orientation might affect tax avoidance. The most likely area where this could prevail is arguably the perceived costs of tax avoidance. Since tax avoidance is a largely hidden practice, that is, the market is not aware of the extent to which it occurs at individual firms, the benefits of lower taxes will not immediately be reflected completely in share prices., but the higher profits that follow necessarily will. This reasoning of hidden practices holds even stronger for the subjective expected costs of tax avoidance, which are allowed to vary by shareholder. These costs will presumably increase over time and with the extent of tax avoidance that the firm adopts. Therefore, the benefits of tax avoidance will present themselves to short- and long-term shareholders in the form of a
8Unfortunately, this line of thought cannot be validated empirically in this paper, since all firms in the sample are publicly traded firms.
9Each firm therefore draws an income Y
tfrom f(.)in each period, ensuring that income varies between periods, but on average (in expectation), each firm will earnR
Ytf(Yt)dYt.
higher share price, but the expected costs will be substantially higher for long-term shareholders, since they will have a higher probability of being a shareholder when these costs materialize.
To account for this dimension of costs, a probability of detection φ(α)is included.
Assumption 1. The probability φ(α)of detection is decreasing and convex in the degree of tax avoidance α: φ0(α) <0, φ00(α) >011. Furthermore, φ0(0) =0 and φ(.) ∈ [0, 1].
Since this is a probability of incurring costs, denote with C the actual fixed cost that is incurred with probability φ(α)12. As discussed, these costs can be composed of a number of items,
presumably reputational costs, fines and voluntary payments represent the largest components. The main point here is that these costs decrease firm value in expectation, and more so if they are incurred with a larger probability, that is, by avoiding more taxes.
The core point of the model is made by making either φ(α) or C variable over time. Two
general approaches can be taken for this. First, an explicit discount factor for the time elapsed since time t=0 can be included in the model. Then, adding a similar expression which discounts the expected costs of tax avoidance as the short-term orientation of shareholders increases yields a full expression of firm value where the expected costs of tax avoidance vary over time and with the time preferences of the shareholder base. However, this approach explicitly models (by adopting a particular functional form) how time and time preferences affect the expected costs of tax avoidance. This is an ambiguous approach, since we cannot know the explicit functional form. A second, more general, approach is therefore to add parameters for time and investor time preferences to the φ(.)function, and only specify how this function behaves with respect to these parameters. This latter option is chosen here.
This approach requires two additional assumptions on the φ(.)function. First, the probability of detection is assumed to increase over time. This implies that as firms evade taxes for a prolonged period, it becomes more likely that their activities are discovered. The rate at which the probability of detection increases over time is termed ρ, with ρ≥0.
Assumption 2. The probability of detection φ(.)depends positively on ρ at a decreasing rate. That is, φ is increasing and concave w.r.t. ρ: ∂φ(ρ∂ρ,α) >0, ∂2φ(ρ,α)
∂ρ2 <0. Furthermore, φ(0, α) =0.
Second, to model short-termism an explicit investor time dimension is included. Following Polk and Sapienza (2009), this is modelled as the arrival of a liquidity shock for each investor j. Liquidity shocks are assumed to follow a Poisson distribution, with a mean arrival rate of qj
for shareholder j, and qj∈ [0,∞). Note that a higher qj implies that this shareholder experiences
liquidity shocks more often, and hence he or she wishes to sell shares more often. Therefore, a shareholder with a high qjis necessarily a short-term shareholder13.
Assumption 3. The probability of detection φ(.)depends negatively on investor time preferences qj at a
decreasing rate. That is, φ is decreasing and convex w.r.t. qj:
∂φ(qj,ρ,α)
∂qj <0,
∂2φ(qj,ρ,α)
∂q2j >0. Furthermore, φ(0, ρ, α) =0.
11Note that this follows from the definition of tax avoidance. With α=1, there is no tax avoidance, whilst there is maximal tax avoidance with α=0. Hence, φ(α)is decreasing. Convexity is assumed because this probably provides the
best reflection of the nature of tax avoidance, from zero to some tax avoidance increases the probability of detection more than from 90 to 100 percent tax avoidance.
12Naturally, it is realistic to assume that C depends on α as well. However, this does not alter the insights derived from the model and therefore C is simply presented as a lump sum fixed cost.
From the above, a complete description of firm value after-taxes can be obtained. In each period t equation (1) reflects the value obtained from income after taxes and managerial diversion. Calculating the discounted value of this stream of cash flows and subtracting the costs associated with it yields total firm value at t=0 as perceived by shareholder j:
Vj0= Z ∞ 0 (1−ατ)β Z Ytf(Yt) dYt−φ(qj, ρ, α)C e−rtdt, (2)
with r the discount rate. Note that the word perceived is stressed, because the probability of detection is not defined as the objective probability that tax avoidance is detected. Rather, the fear a shareholder has for negative consequences from tax avoidance dominates how qj and ρ
are determined. Therefore, the probability of detection is a subjective probability. To arrive at a relatively objective measure of firm value, it is assumed that managers take the preferences of the average shareholder into consideration when determining what the preferred time orientation for the firm should be (Stein, 1989; Polk and Sapienza, 2009; Derrien et al. 2013). Effectively, this implies that the subscripts for a particular investor (j) should be dropped.
From here on, we can assume that managers rationally maximize the average shareholder’s wealth in a straightforward way, with no agency problems occurring. The optimization problem is then relatively simple and consists of an appropriate choice of α (since β is necessarily 1 in case of such an ideal agent) in equation (2). Alternatively, we can assume self-interested managers, who maximize their own pay. This is the approach that Desai and Dharmapala (2006) follow, by creating an additive function of several components of compensation that the manager optimizes. This is also the approach that is followed here.
The first component of compensation consists of the utility managers derive from diverting resources. From equation (1), managers divert an amount of(1−β)Y¯t. The manager attaches
utility to his income in line with his utility function w(.).
Assumption 4. The manager’s utility function w(.)is increasing and concave in its arguments: w0(.) > 0, w00(.) <0. Furthermore, w(0) =0.
In addition, the manager is compensated with equity-based incentives to assure that the manager puts in effort and does not divert all the income generated. This is modelled by denoting the manager’s sensitivity to firm value by θ14. Hence, the second term in the manager’s compensation is θV0. Finally, there will be costs associated with managerial diversion. These costs
could be in the form of effort to divert or cover up the diversion, or costs that occur upon discovery. They depend on the level of β, with a high β reflecting low costs, since there is low managerial diversion in this case. Correspondingly, a low beta implies large managerial diversion and hence there will be high costs associated with it. Moreover, firms with short-term oriented shareholder bases tend to have poorer corporate governance structures (Harford et al., 2015). Therefore, these firms allow for greater rent extraction. One line of argumentation explaining why this happens is the lack of incentive to monitor management when an investor is only interested in holding shares for a short period of time. Thus, the costs of diversion decrease as q increases at a decreasing rate, in other words, higher short-termism reduces the costs of managerial diversion because there is less monitoring by shareholders. These characteristics are reflected in the following assumption:
Assumption 5. The manager’s cost function ψ(.)is decreasing and convex in its arguments: ψ0(.) < 0, ψ00(.) >0. Furthermore, ψ(1) =0.
In sum, the manager will thus maximize the following expression: max
α,β
U0=w((1−β)Yt) +θV0−ψ(β, q), (3)
with V0defined by equation (2) (after dropping the subscript j). Note that there are no contracting
issues as in standard principal-agent models (Holmström, 1979) and this class of models in tax avoidance research (Chen and Chu, 2005; Crocker and Slemrod, 2005; Biswas et al. 2013). Rather, as discussed before, the approach of Desai and Dharmapala (2006) is employed to arrive at simple predictions of the relation between tax avoidance and short-termism. To finalize the discussion of the manager’s optimization problem, there is one constraint that has not yet been discussed. V0≥0, since negative market values are impossible.
Empirical predictions of the model
In order to evaluate whether short-termism affects tax avoidance, the optimization problem stated in equation (3) is solved. The resulting implicit solution for α will allow for a comparative static analysis which reveals the effects of a small increase in the parameters on the endogenous variables. Of the two endogenous variables, the level of tax avoidance α is the most important one. However, the predictions for β when the parameters change are also derived.
Inserting equation (2) into equation (3) yields the full program of the manager15: max α,β U0=w((1−β)Yt) +θ Z ∞ 0 (1−ατ)β Z Ytf(Yt) dYt−φ(q, ρ, α)C e−rtdt −ψ(β, q), (4) subject to the constraints on the variables that were previously discussed. Solving this maxi-mization problem, assuming that the first-order condition (FOC) approach is valid, yields two FOCs: ∂U0 ∂α = θ r −τβ ¯Yt−φα(q, ρ, α) =0, (5) ∂U0 ∂β = −Ytwβ((1−β) ¯ Yt) +θ r (1−ατ)Y¯t−ψβ(β, q) =0. (6)
These FOCs implicitly define the optimal level of α and β. To obtain the comparative static effects of small changes in q and θ,we can differentiate the FOCs with respect to these parameters in turn16. First, consider the comparative static analysis of small changes in q:
"∂α∗ ∂q ∂β∗ ∂q # = −θ rφαα(., ., .) −θrτ ¯Yt −θ rτ ¯Yt Yt2wββ(.) −ψββ(., .) −1− φαq(., ., .) −ψβq(., .) .
Since we are interested in the sign of the derivatives with respect to q only, and we have posited that the determinant of the Hessian matrix is negative, we can use the following formulation based on Cramer’s rule, which allows us to draw conclusions about signs of the comparative static effects. Note that there is an additional minus in front of the sign, to account for the effect of
15After stating the managers full program, the fixed costs C are ignored, since they are constant and will not affect any of the outcomes.
16In order to assure that a maximum is found, the determinant of the second-order conditions should be negative (the Hessian matrix should be negative (semi)definite). This holds as long asθ
r τ ¯Yt 2
the negative determinant (which is not explicitly included since only its sign is interesting in this situation) on the sign of these derivatives:
sign ∂α∗ ∂q = −sign −φαq(., ., .)Yt2wββ(.) −ψββ(., .) − θ rτ ¯Ytψβq(., .) , (7) sign ∂β∗ ∂q = −sign θ r φαα(., ., .)ψβq−τ ¯Ytφαq(., ., .) . (8)
Given the expressions for the signs of these derivatives, a condition can be formulated to arrive at empirical predictions regarding the effect of short-termism on tax avoidance. The condition ensures that short-termism increases tax avoidance and managerial diversion. The subsequent empirical tests will verify whether this condition and its corresponding proposition will hold. The condition (suppressing arguments) is:
Condition 1. φαq<min φααψβq τ ¯Yt , − ψβqθrτ ¯Yt Y2 twββ−ψββ
Depending on ψβq, both arguments on the right-hand side are either negative or positive. Combining this condition and equation (7), we can formally state the relationship between short-termism and tax avoidance as follows:
Proposition 1. Conditional on the second-order conditions being satisfied, and the validity of assumptions
1 through 5, a small increase in short-termism results in an increase in corporate tax avoidance, if condition 1 is satisfied. If condition 1 does not hold, the converse will occur: a small increase in short-termism will decrease corporate tax avoidance. Formally, under condition 1: ∂α∗
∂q < 0, with α
∗implicitly defined by
equation (5).
Proof. Follows directly from the assumptions, condition 1 and equation (7).
Similarly, a corresponding proposition can be drafted for managerial diversion, since condition 1 also ensures a negative coefficient for this derivative:
Proposition 2. Conditional on the second-order conditions being satisfied, and the validity of assumptions
1 through 5, a small increase in short-termism results in an increase in managerial diversion, if condition 1 is satisfied. If condition 1 does not hold, the converse will occur: a small increase in short-termism will decrease managerial diversion. Formally, under condition 1: ∂β∂q∗ <0, with β∗implicitly defined by equation (6).
Proof. Follows directly from the assumptions, condition 1 and equation (8).
To finalize this section, a similar comparative static analysis for small changes in θ, the degree of exposure to equity incentives, is performed. Starting from the derivatives in equations (5) and (6), we can reformulate the system of two linear equations:
"∂α∗ ∂q ∂β∗ ∂q # = −θ rφαα(., ., .) −θrτ ¯Yt −θ rτ ¯Yt Yt2wββ(.) −ψββ(., .) −11 r −τβ ¯Yt−φα(., ., .) 1 r (1−ατ)Y¯t . Again, using the partial Cramer’s rule to determine signs yields (suppressing arguments):
sign ∂α∗ ∂θ = −sign 1 r (τβ ¯Yt+φα)(ψββ−Yt2wββ) + θ r2(1−ατ)τ ¯Yt 2 , (9) sign ∂β∗ ∂θ = −sign θ r2 φαα(1−ατ)Y¯t + −τβ ¯Yt−φα τ ¯Yt . (10)
Condition 2. τβ ¯Yt≥ −φα.
From this condition and equations (9) and (10), two propositions can again be derived on the relationship between equity incentives, tax avoidance and managerial diversion. First, the model shows that equity incentives are related to tax avoidance as follows:
Proposition 3. Conditional on the second-order conditions being satisfied, and the validity of assumptions
1 through 5, a small increase in equity incentives results in an increase in tax avoidance, if condition 2 is satisfied. If condition 2 does not hold, the converse will occur: a small increase in equity incentives will decrease tax avoidance. Formally, under condition 2: ∂α∗
∂θ <0, with α
∗implicitly defined by equation (5).
Proof. Follows directly from the assumptions, condition 2 and equations (9).
Second, the relationship between equity incentives and managerial diversion is as follows:
Proposition 4. Conditional on the second-order conditions being satisfied, and the validity of assumptions
1 through 5, a small increase in equity incentives results in a decrease in managerial diversion, if condition 2 is satisfied. If condition 2 does not hold, the converse will occur: a small increase in equity incentives will increase managerial diversion. Formally, under condition 2: ∂β∗
∂θ >0, with β
∗implicitly defined by
equation (6).
Proof. Follows directly from the assumptions, condition 2 and equation (10).
This concludes the discussion of the model. All propositions are in principle empirically testable, although the primary focus is on the relationship between short-termism and tax avoid-ance, and to a lesser extent this study attempts to confirm the evidence on equity incentives and tax avoidance, in line with previous studies (Phillips, 2003; Minnick and Noga, 2010; Rego and Wilson, 2012; Armstrong et al., 2015). The additional propositions on managerial diversion are a consequence of the modelling approach followed (Desai and Dharmapala, 2006), but are nevertheless interesting to re-examine empirically within this dataset.
IV.
Methodology
Data
To test the predictions of the model presented in the previous section, a sample must be selected. The most difficult data to collect is probably a measure of short-termism. Generally, two methods can be employed to measure short-termism. First, measuring symptoms or potential causes of short-termism is an often employed empirical strategy in research on short-termism. For example, Roychowdhury (2006) estimates discretionary R&D expenditures to capture real activities management. Arguably, real activities management where long-term investments in activities such as R&D, advertising and capital expenditures are reduced, are an important tool for achieving short-term profits. Alternatively, investor time orientation is often used, where long holding periods or the mere fact of being an institutional investor is often assumed to be an indication of a long-term orientation (Derrien et al., 2013; Harford et al., 2015).
Table 1: Conference calls
Panel A: Sample selection & processing
Initial number of quarterly conference calls 82,101
Calls for which the presentation and Q&A sessions are inseparable 588 Calls from firms who did not have a fiscal-year end month in Compustat 990
Calls from firms who are not in the Compustat database 22,947
Quarterly calls linked to Compustat identifiers 57,576
Calls lost due to aggregation within quarters 5,908
Calls lost due to aggregating over quarters 35,114
Firm-years after summing word counts over quarters 16,554
Panel B: Number of aggregated word counts from conference calls per year
2004 1 2005 183 2006 271 2007 1,564 2008 1,927 2009 1,595 2010 1,340 2011 1,208 2012 2,239 2013 2,746 2014 3,059 2015 421 Total 16,554
are explored to retain a broad perspective on the potential influences of short-termism on tax avoidance. This has consequences for the sample selection procedure, since earnings conference calls are required in order to obtain estimates of short-termism following the Brochet et al. (2015) and DesJardine and Bansal (2015) approaches.
The sample is thus restricted to firms that have published earnings conference call transcripts on seekingalpha.com17. Note that there are not always conference calls available for each year,
however, data on financial variables might be more readily available. Hence, the number of observations for financial variables might be substantially higher in some cases. The constant factor throughout the dataset is the number of firms, all data is collected form the same set of firms, which is based on those firms who have earnings conference calls on seekingalpha.com. The transcripts of conference calls are usually on a quarterly (or semi-annual) basis, thus yielding multiple calls per year, at occasion even multiple per quarter. In sum, 82,101 quarterly conference calls are available from seekingalpha.com. Table 1 breaks this initial number down into the final number of firm-years that will yield a measure of short-termism. First, the calls are separated into the presentation part, and the Q&A part of the conference call. The former includes a presentation by management on this quarter’s results, and is therefore largely scripted and its contents are
predetermined. The Q&A part consists of sell-side analysts who are allowed to ask questions to management. As such, this part of the conference call is not scripted.
Approximately 588 calls could not be separated into a presentation and a Q&A session. Randomly inspecting some of these calls warrants the conclusion that these calls are a) not a quarterly earnings conference call or b) they skipped either one of the two sessions: only a presentation was held, or alternatively only a Q&A session was conducted. Since it was not possible to separate a) from b), all these calls were dropped altogether. Next, the calls contain ticker symbols, which are used to link them to the Compustat North-America database in order to link a fiscal year to each conference call. A total of 990 conference calls did not contain a fiscal-year end month in Compustat. Therefore, it was not possible to identify which quarter these conference calls belong to, and hence these 990 calls are also deleted from the sample. Next, 22,947 conference calls could not be linked to the Compustat North-America database. These firms are likely foreign, and because other data would be difficult to collect for these firms, they are also removed from the sample. A total of 57,576 quarterly conference calls remain, for which at least a ticker symbol and a fiscal year-end month is available in Compustat.
Next, words are counted according to the word lists used by DesJardine and Bansal (2015) and Brochet et al. (2015). Subsequently, these word counts are summed in case there are multiple calls for a single quarter, after which only the summed word count is retained for each quarter. This further reduces the sample with 5,908 conference calls. Finally, for the main analysis an annual average of short-termism is used, thus the word counts are summed over all four quarters (or less, if there are fewer quarters for which a conference call is available). This yields a final set of firm-year word counts of 16,554. Panel B of table 1 breaks the total number of firm-years down over the time period that is examined, in order to assess how the sample develops over time. As could be expected, conference calls in early years are quite scarce, whilst coverage or availability increases steadily in later years.
From this sample of conference calls, the ticker symbols are used to extract data from Com-pustat, Execucomp, BoardEx and the MSCI GMI ratings. The time period used is substantially increased to allow for a substantial number of lagged values, which is helpful when alternative measurements are employed and a system GMM approach is implemented to account for endo-geneity. From the merged dataset, those firms who have no or very little (<$10,000) total assets are dropped. In addition, those observations that have no data on either one of the five measures of tax avoidance are deleted as well. For the regressions including all variables, data on each of the variables must be available from all sources. Often, this is not the case, which strongly reduces the number of observations available for individual regressions. Moreover, this dataset includes years from 2002 till 2015, which is necessary to calculate average effective tax rates and the lagged structure for the system GMM estimates.
Measurement
Measuring tax avoidance
and it does so on each of five different aspects of tax avoidance. Before turning to the overall measure, the individual components are discussed.
The most visible18measure of tax avoidance in firm’s disclosures is the effective tax rate (ETR). This measure is commonly used as an approximation of tax avoidance, with a lower ETR implying less taxes being paid (Dyreng et al., 2008; Minnick and Noga, 2010; Rego and Wilson, 2012; Armstrong et al., 2015; Huseynov, Sardarli and Zhang, 2017). Earlier research often concentrated on the accounting definition of the ETR, total income taxes over the year, divided by pre-tax income. Dyreng et al. (2008) argued that this measure is not appropriate for its intended purposes, and they introduced the long-run cash ETR:
CETRi,t+T= ∑
T
t Taxes Paidit
∑T
t Pre-tax Incomeit−∑Tt Special Itemsit
,
with T commonly chosen to be 3 or 5 years, and t+T is the current time period. In other words, this is the T years historical cash ETR (CETR). The summing over years ensures that one-time cases such as an incidental loss are filtered out of the ETR, to provide a clearer picture of the tax burden the firm actually faces. A similar measure is calculated for the GAAP ETR (GETR), and the choice of T is 3 years in both cases19. Finally, for both measures the industry median is subtracted, since ETRs vary greatly among industries, but less so within industries.
Second, a measure of tax avoidance that was recently developed by Dyreng and Lindsey (2009) uses data from annual SEC filings to extract where companies report to have material subsidiaries. Using widely acknowledged lists of countries that are reported to be a tax haven, the authors were able to determine which firms had subsidiaries in tax havens. Subsequent research has employed this (or comparable) data20 to construct measures of tax avoidance in a variety of contexts and applications (Dyreng, Hanlon, and Maydew 2012; Dyreng, Lindsey and Thornock, 2013; Black, Dikolli and Dyreng, 2014; Akamah, Hope and Thomas, 2017). Two variants are constructed here to arrive at an explicit measure:
PPT_SUBSit= # o f subsidiaries in tax havenit
# o f subsidiariesit
,
PPT_TXH AVit = # o f tax haven countries where f irms have subsidiariesit
# o f countries where f irms have subsidiariesit
.
Third, the degree of uncertain tax benefits at year-end is used to approximate tax avoidance21. This item reflects the estimates of management regarding positions that will be disputed by the tax authorities if there is a tax audit. This is scaled by total assets (UTB), following Armstrong et al. (2015). According to Lisowsky et al. (2013), the extent of uncertain tax benefits correlates strongly with firms’ actual tax avoidance practices.
Fourth, Wilson (2009) used data on firms that were actually identified as users of tax shelters to estimate how financial characteristics could explain involvement in these shelters. He ran several regressions to identify the probability of engaging in sheltering activities. In subsequent empirical research, the coefficients estimated by Wilson (2009) were used to identify which firms in a sample
18And therefore likely also the one that is most influenced by firm management for window-dressing purposes. 19Robustness checks are also performed with longer CETR and GETR, because researchers tend to use longer time periods to measure the long-run ETR (Dyreng et al., 2008; Minnick and Noga, 2010). However, also a shorter time period to measure the ETR might be interesting in the context of this study. Since the topic is short-termism, it is not very likely that the three-year ETR will reflect adjustments due to short-termism. In any case, it seems more likely that a short-run ETR is affected by short-termism more easily than a long-run ETR. Thus, robustness checks will employ a single year ETR as well.
were likely to engage in tax avoidance (Rego and Wilson, 2012; Boone, Khurana and Raman, 2013). The exact formula is as follows:
ln
P 1−P
= −4.30+6.63BTDit−1.72LEVERAGEit+0.66SIZEit+2.26ROAit+
1.62FOREIGN_I NCOMEit+1.56RDit,
with BTD the book-tax difference, obtained by grossing up taxes paid as described in Wilson (2009), and subtracting this from book income. LEVERAGE is long-term debt divided by total assets. SIZE is the natural logarithm of total assets, and ROA is the net income divided by total assets. Next, FOREIGN_INCOME is a dummy that equals 1 if the firm has foreign income and zero otherwise. Finally, RD is a firm’s R&D expenditures divided by total assets. From the outcome of this regression, rewriting the predicted y to obtain P results in the identification of which firms are predicted to engage in tax sheltering. This is the measure that is used to as a proxy for tax avoidance, and it is termed SHELTER. The previous measures often included the possibility of firms doing honest business with their subsidiaries (PPT_SUBS, PPT_TXHAV), or firms could have a dispute about their tax position with the IRS on whether something is legitimate (UTB). However, this measure attempts to identify those firms that actually engage in illegal tax avoidance.
Finally, the fifth measure is in a sense similar to the previous one, since it is based on an empirical specification, which is used to calculate the desired measure. Frank, Lynch and Rego (2009) extended the method of measuring tax avoidance pioneered by Desai and Dharmapala (2006), namely the degree of book-tax differences. Based on a set of regressors, these authors identified the temporary part of book-tax differences. Most influences on the book-tax difference that are derived from application of tax laws concern permanent differences, thus this method seeks to identify the temporary differences to measure tax avoidance. Again, subsequent empirical research has also used this measure to investigate tax avoidance (Rego and Wilson, 2012; Hasan et al. 2017). To calculate this measure, the specification used by Frank et al. (2009) is employed to obtain the residuals of the following regression by two-digit SIC industry and year:
PERMDIFFit =β0+β1I NTANGit+β2UNCONit+β3MIit+β4CSTEit+β5∆NOLit+
β6PERMDIFFi,t−1+εit,
with PERMDIFFit=BIit− [(CFTEit+CFORit)/STRit] − (DTEit/STRit). The definition of each
of these variables is discussed in appendix A, where all variable definitions are explained. The actual measure, as mentioned, consists of the residuals εit. Again, this measure takes a different
angle to the materialization of tax avoidance. Book-tax differences are an important indicator that firms are actively managing their tax liability, especially when permanent differences are removed, as is the case in this measure.
There are several remarks to be made on these measures. As suggested, the aim was to capture different aspects of tax avoidance. ETRs are the broadest measure, incorporating virtually all decisions that affect the amount of taxes paid. Specifically, they will measure a substantial amount of legal deductions as well22. As such, a measure like an ETR perhaps puts more emphasis on which firms are actively engaged in strictly legal tax management, compared to those who are less so. Compared to an ETR, the UTB measure is an improvement, in the sense that it only captures those tax related matters which might be disputed by the tax authorities. It therefore quite distinctly reflects the degree to which a firm engages in risky tax planning. Similarly, the SHELTER measure attempts to identify those firms that engage in risky tax planning, but goes
one step further by focusing on strictly illegal, instead of debatable, practices. On the other hand, the book-tax measure DTAX takes a step back, and measures a symptom of tax avoidance, namely large permanent book-tax differences for those firms that are heavily engaged in tax avoidance.
At first glance, the presence in tax havens or the identification of firms using tax shelters might be the best measure to capture the widely criticized tax avoidance practices by multinational firms. These measures neatly describe to what extent a firm has set-up a network for the sole purpose of tax avoidance. However, the tax haven measure, with a few exceptions, also includes countries such as Switzerland and the Netherlands as a tax haven. Whether this is true or not, it is also perfectly feasible that the firm has actual business operations in such countries, such that tax considerations might be a secondary concern. Other countries listed as tax havens do not share this problem, since they often have virtually no profit tax and a very small customer base due to the low number of inhabitants; in such cases it seems likely that tax considerations were the primary reason to register a subsidiary in this country. Moreover, the SHELTER measure has the downside that it is rather simplistic to assume it is possible to identify shelter participants based on six characteristics from the firm’s financial statements.
In sum, all measures have their advantages and their disadvantages, which illustrates why research on corporate tax avoidance should consider several measures instead of one. In an attempt to capture all these measures in one distinct metric, a z-score is calculated for each of the metrics. Thus, each observation then reflects the number of standard deviations this particular observation is away from the mean. After correcting the ETR measures (for which a positive higher score reflects less tax avoidance instead of more, as is the case for the other measures), the z-scores are summed and the average is calculated. Missing values are not included, so it is possible that the measure is calculated over for example only six or five measures instead of seven. The index is termed TX_IND, and a higher value reflects higher tax avoidance.
Measuring short-termism - Textual analysis
As discussed, two approaches to measuring short-termism are employed. The first is an application of the dictionary approach to content analysis. This entails the use of word lists that suppos-edly measure a concept. Recent reviews of the literature that discuss this and more advanced approaches to content analysis are Li (2010), Kearney and Liu (2014) and Loughran and McDonald (2016). Examples of empirical studies that use textual analysis to measure concepts in a business environment are abundant: Larcker and Zakolyukina (2012) use it to measure whether CEOs and CFOs use deceptive language in conference calls, Huang, Zang and Zheng (2014) show that analyst reports contain important information to investors over and above other sources of information, Merkley (2014) examines R&D disclosures and relates it to earnings performance, Shin and You (2017) examine whether CEOs can influence their pay by talking about shareholder value in firm disclosures. Finally, Lo, Ramos and Rogo (2017) link annual report readability to earnings management.
then any call in January, February or March of the next calendar year is identified as a conference call from the previous year’s Q4 quarter, since the contents of the call discuss the final quarter of that fiscal year. After this, all the conference calls are now linked to a particular fiscal year that corresponds to the right fiscal year on which Compustat holds data. Although there might be some mismatches in exceptional cases, this strategy of assigning fiscal years should deliver correct results in the large majority of cases. Random checks of conference call content and the assigned fiscal year confirms this expectation: no wrong assignments of the fiscal year are found. The same procedure was applied to other quarters: conference calls in the three months after the end of a quarter are assigned to be a call on the most recently ended quarter.
Then, to calculate the measure, words were counted along two dimensions: one being the long-term oriented words, the other its converse, words indicating a short-long-term orientation. Appendix B contains a table (table 9) listing the respective word lists developed and used by Brochet et al. (2015) (panel A) and DesJardine and Bansal (2015) (panel B) in constructing their measures. These word lists are both employed to count words in the 57,576 quarterly conference calls mentioned in panel A of table 1. As described, these word counts are then aggregated into an annual sum of word counts23. The respective measures are then calculated as follows:
Short-termism_Bit =
Short-term wordsbit
Long-term wordsbit
, Short-termism_Dit =
Long-term wordsdit
Long-term wordsdit+Short-term wordsdit,
with the subscript b and d an indication of which word list was used. The addition B or D indicates which measure, either Brochet or DesJardine, is represented with this variable.
Measuring short-termism - Real activities
In order to derive other measures that could possibly measure short-termism, two options were available. Either antecedents or consequences of short-termism could be used. Past research has often chosen to focus on antecedents of short-termism, such as analyst coverage or long-term investors, this study, however, uses discretionary real activities management as an alternative to the textual measures of short-termism. Real activities management is necessarily shifting the focus towards the nearer future if managers decide to decrease, for example, R&D expenditures, or capital investments, simply because there are less resources dedicated to creating value in the future. Therefore, real activities management captures the degree to which a firm is able to deliver future profitability, and it is thus a measure that closely resembles short-termism.
Specifically, discretionary R&D expenditures and similarly calculated discretionary capital expenditures are the measures of real activities management which are used to reflect short-termism. The design developed by Roychowdhury (2006) is implemented to estimate these measures. This amounts to running the following regression by industry (Fama-French 48 industry classification) and year:
R&D expensesit
Total assetsi,t−1
=α+β1
1 Total assetsi,t−1
+β2
Salesi,t−1
Total assetsi,t−1
+εi,t,
and similarly for Capital expenditures. For each actual observation the predicted value based on this regression is subtracted, yielding the error term of this regression for each observation. This is
the measure of real activities management. A higher measure implies a more long-term oriented focus since the firm spends more than predicted on the respective indicator of long-term oriented investment.
Measuring governance
Under the heading of governance, two variables are discussed. First, an index is created along the lines of Aggarwal et al. (2009). This measure of governance is used both as a control variable and as an approximation of managerial diversion, since a strong governance environment generally leaves less scope for managerial diversion of resources. Second, to test the comparative static effects, a measure of equity incentives is necessary. For this, the approach by Cheng, Hong and Scheinkman (2015) is employed. Brochet et al. (2015) also use this measure to identify whether higher equity incentives increase short-termism24. The measure is the residual of a regression, specifically, the average stock-based compensation of the firm’s executives on market capitalization, market-to-book ratio, a constant and industry and year fixed-effects. The measure is termed EQUITY_INC.
Returning to the governance index, Aggarwal et al. (2009) use a list of 44 items from ISS to formulate certain rules of good governance. If firms satisfy this rule, the item is assigned a 1, otherwise it is 0. After checking for all 44 items, the items are summed and divided by the number of non-missing items, to arrive at a proportion of items that are satisfied. This constitutes their measure of good governance. The approach taken here is similar, although in a slightly reduced form. From the MSCI GMI Ratings, a substantial amount of information on boards, anti-takeover measures, compensation and executives is obtained. Using the list of Aggarwal et al. (2009), an attempt is made to derive similar or identical rules from the MSCI data. This eventually results in a total of 26 items (for a complete list, see appendix C, table 10). By following the procedure described above, the measure GOV_INDEX is obtained as an approximation for the quality of the internal governance environment. For robustness checks, two board characteristics that Armstrong et al. (2015) used to approximate the internal governance environment, the proportion of independent directors and the number of financial experts, are also included. These variables are defined as independent directors divided by total directors (PPT_IND) and simply the number of financial experts (NUM_FIN).
Measuring control variables
In addition to the variables of explicit interest in this study treated till this point, there is also a substantial number of control variables that are commonly employed in research on tax avoidance. First, the size of the firm is included by taking the natural logarithm of the firm’s market value (SIZE)(Minnick and Noga, 2010; Armstrong et al., 2015). Second, the cash flow from operations, scaled by total assets, is included (Armstrong et al., 2015). Third, the book-to-market value of equity is included, to account for firms that are over- or undervalued. Overvalued firms might have a stronger incentive to manage their tax burdens downwards, to keep earnings high and prevent a drop in their stock prices (Minnick and Noga, 2010). Fourth, a measure of leverage, by means of the debt-to-equity ratio (DE) is important, since higher levered firms are able to use the interest tax shield more to a greater extent25. Fifth, the return on assets (ROA) is included as a
24In the tax avoidance literature, the Core and Guay (2002) measure is commonly used to examine to what extent managers will react to higher equity sensitivity. However, a lack of time unfortunately rendered the calculation of this measure infeasible. Therefore, I resorted to the arguably less preferable but more easily obtainable measure of Cheng et al. (2015)
measure of profitability. More profitable firms generally have more to hide from the tax authorities (Minnick and Noga, 2010; Rego and Wilson, 2012; Hasan et al., 2017).
Sixth, in some regressions (if the variable is not used for other purposes) a measure of R&D or capital expenditures can be included, since such expenditures are often tax subsidized. Seventh, two measures accounting for losses which might have tax consequences are included. One is a dummy variable that measures whether the firm has suffered a loss last year (LOSS), the other is a dummy measuring whether the firm reports any tax loss carry forwards (NOL)(Minnick and Noga, 2010; Rego and Wilson, 2012; Boone et al., 2013; Hasan et al., 2017). Eighth, two measures of how asset intensive firms are might be important. Intangible assets divided by total assets measures to what extent firms hold intangibles (INTANG), who often have special tax treatments. Property, plant and equipment is generally one of the largest items on which firms write off, thus firms with more physical capital tend to have larger depreciation, and hence this might affect (temporary) book-tax differences and ETRs (Boone et al., 2013; Hasan et al., 2017). To obtain the measure it is again divided by total assets (PPE). Finally, firms with foreign income face different tax regimes and are more likely to have foreign tax structures in place. Hence, a dummy that equals 1 if the firm earns foreign income is included (FOR_INC) (Boone et al., 2013; Hasan et al., 2017).
Empirical specification
The baseline empirical specification employed in all regressions is as follows26:
Tax Avoidanceit =γ0+γ1Short_Termismit+γ2 EQU ITY_I NCit+γ3GOV_I NDEXit+ J
∑
j=4
γjControl variablejit+Year f ixed e f f ectst+εit, (11)
where the set of control variables J varies depending on the other variables included. For example, the first control variable j= 4 is thus SIZE, the second j= 5 is CFOps, etcetera. Furthermore, Tax Avoidance can be one of the eight variables that measure tax avoidance, but mostly it refers to TX_IND, the composite measure of tax avoidance. Similarly, Short_Termism can refer to either one of the two textual measures (even the subsets of the conference calls analysed, presentation and Q&A sessions are separate variables in robustness checks, so in total there are six textual measures of short-termism), or the two real measures: discretionary R&D expenditures or discretionary capital expenditures. A final twist is used to verify whether propositions 2 and 4 hold, by making GOV_INDEX the dependent variable, whilst disregarding tax avoidance altogether.
The main estimation results are obtained with the straightforward fixed effects within estimator. Since the dataset used for the empirical results consists of a multi-year panel with more individuals (firms) than time periods, this is the appropriate estimator for a first glance at the validity of the hypotheses specified in the propositions in section III. Furthermore, the work of Armstrong et al. (2015) revealed that equity incentives and governance have different effects on tax avoidance along its distribution. That is, variables might affect tax avoidance vastly different at different points in the distribution, so high tax avoiders might respond different to a strong governance environment than low tax avoiding firms. To assess whether the model outlined in section III holds independent of the level of tax avoidance, quantile regression27 is used to assess whether
26Note that the set of control variables is allowed to vary depending on which dependent variable is employed. Furthermore, the GMM estimates also contain two lags of the dependent variable as regressors.
short-termism has different influences at different points of the tax avoidance distribution. Finally, there are several potential issues with the fixed effects estimator, of which endogeneity is the most relevant and most important issue to address. However, as for example Armstrong et al. (2015) and Desai and Dharmapala (2006) acknowledge, it is prohibitively difficult to find a suitable instrument that would allow for an instrumental variables estimator in order to counter these endogeneity issues. As is often the case in the areas of governance and firms, exogeneity of the instrument cannot be ensured. However, Wintoki, Linck and Netter (2012) suggest an alternative approach to counter endogeneity problems in empirical research on the internal governance situation of firms: system GMM for dynamic panels might be an appropriate method to solve these endogeneity issues. With system GMM, the instruments are created from within the dataset, by using lagged values of the variables to generate instruments. By doing so, it is possible to solve the endogeneity issue, and make causal inferences. Although this is not the main goal of this study, additional analysis will apply system GMM in a similar fashion as Minnick and Noga (2010) did, along the lines of the manual by Roodman (2009), with the Windmeijer (2005) correction to counter the bias in standard errors when the two-step approach is used.
Table 2: Time orientation of firms and industries.
Panel A: Industry classification
Long-term oriented industries Short-term oriented industries
Beer & Liquor Banking
Defense Electronic Equipment
Agriculture Petroleum and Natural Gas
Shipping Containers Measuring and Control Equipment
Aircraft Pharmaceutical Products
Consumer Goods Steel Works
Non-Metallic and Industrial Metal Mining Shipbuilding, Railroad Equipment
Textiles Trading
Recreation Computers
Rubber and Plastic Products Medical Equipment
Panel B: Firm classification
Long-term oriented firms Short-term oriented firms
Sony Corp. Blackrock Capital Investment Corp.
Electronic Arts Inc. TICC Capital Corp.
General Mills Inc. Golub Capital BDC Inc.
Coca-Cola European Partners Hutchinson Technologies Inc.
United Technologies Corp. Provident Financial Services Inc.
Goodyear Tire & Rubber Co. Fulton Financial Corp.
Whirlpool Corp. Vanda Pharmaceuticals Inc.
Time Warner Inc. Applied Micro Circuits Corp.
PepsiCo Inc. Blackberry Ltd.
