Estimating Downside Tax Risk:
Exploration of Unfavorable Tax Settlements
Andrew M. Bauer∗ Department of Accountancy
The University of Illinois at Urbana-Champaign
Kenneth J. Klassen School of Accounting and Finance
University of Waterloo
This draft: April 2014
Keywords: tax risk, risk management, tax settlement, extreme downside risk
*Corresponding author: Andrew M. Bauer, Department of Accountancy, The University of Illinois at Urbana-Champaign, 1206 S. Sixth Street, Champaign, IL 61822, tel: +1-217-244-4505, email: [email protected]. We appreciate the comments received by Katharine Drake, Tom Omer and Kathleen Powers. We also appreciate the comments from workshop and/or conference participants at Northeastern University, the University of Tennessee, the University of Texas-Austin, the University of Utah and the 2014 ATA Mid-Year Meeting. We thank Qin Cao, Christina Chen, Armando Espinal and Diana Rechenmacher for excellent research assistance. We gratefully acknowledge the financial support provided by the University of Illinois at Urbana-Champaign.
mailto:[email protected]
Estimating Downside Tax Risk: Exploration of Unfavorable Tax Settlements
Abstract
In this paper, we examine downside tax risk as the likelihood that a firm will face a large and negative tax outcome. To achieve this objective, we develop a model to predict the probability that a firm will have an unfavorable settlement with tax authorities, a “tax loss event.” We combine an industry-adjusted measure of high cash effective tax rate (Cash ETR) values with financial statement verification of a tax settlement to identify a sample of tax loss event firms. We develop a logistic model from a parsimonious set of factors that are associated with downside tax risk to generate a scaled probability of the likelihood of these negative events, the DTR-score. Empirical tests show that our model provides strong predictions of a tax loss event. We further consider whether market participants price this risk and the timing of any reaction. We show that the abnormal returns are negative in the event year for firms with a tax loss event and non-positive in the prior year as well. These results do not appear to vary with the volatility of ETR, a proxy for tax risk management. Our research adds to the emerging literature that examines the tax risk within firms. Specifically, we develop a prediction model useful for various firm stakeholders, as well as provide additional insight into the factors that create downside risk for the firm and how the market reacts to the revelation of such risks.
1. Introduction
In recent years, the concept of tax risk within corporations has received much attention.
The Organization for Economic Cooperation and Development (OECD), as well as major
accounting firms, highlight the links between tax risk and uncertainty, and publicize the role of
tax risk management within strong corporate governance structures (PricewaterhouseCoopers,
2004; OECD, 2009). Corporate managers and boards of directors have begun to recognize tax
risk management as a separate component of enterprise risk management (Wunder, 2009). Yet,
what does it mean to say that a firm has higher tax risk? According to Neubig and Sangha
(2004), tax risk is the possibility that the firm will have an unexpected tax outcome. This
definition implies that tax risk involves both unexpected positive and negative outcomes.
However in this paper, we focus our analysis of tax risk on loss events (i.e., events in the extreme
negative tail of a distribution) and examine whether these events are identifiable ex-ante. We
further examine whether the equity market reacts to these events, our proxy for downside tax
risk.
Examining these questions is important for four reasons. First, despite the recent attention
on tax risk, useful, public information about firms’ tax risk is sparse. Hanlon (2003) and Lenter
et al. (2003) acknowledge that although tax disclosures in financial statements are meant to be
informative, they are often lacking. Gleason and Mills (2002) find that U.S. firms infrequently
disclose tax contingencies despite the substantial future payments tied to those contingencies. In
a more recent study, Raedy et al. (2011) conclude that investors fail to use the information
provided in tax footnotes due to a general lack of informativeness. Thus, stakeholders would
benefit from a method to identify tax risk. Second, the question of whether tax risk can be
assessed separately from overall firm risk is unclear. Gleason and Mills (2002) describe the
1
current state of disclosure around negative tax outcomes as one which lacks value-relevant
information for investors. If tax risk management is a component of enterprise risk management
(Wunder, 2009) then investor assessments of firm risk could subsume value-relevant assessments
of tax risk. Our model allows us to create a proxy for downside tax risk ex-ante and also to
examine, if stakeholders react to such information, whether stakeholders value tax risk separately
from other risk factors. Third, prior literature documents that the distribution of risk is
asymmetric and skewed to the downside (e.g. Ang et al., 2006; Huang et al. 2012), thus our
focus on unfavorable tax settlements is consistent with this view. Fourth, tax risk management is
a related concept that the extant literature has not explored. We provide a first step by
investigating whether some managers adjust their GAAP effective tax rates (ETRs) in
anticipation of potential tax loss events.
We address our research questions in the following ways. First, we conceptualize
downside tax risk as the probability that a firm suffers a large tax loss – i.e., an unfavorable
settlement with a tax authority. This concept of tax risk complements existing research that
views tax risk as a form of volatility (e.g., Guenther et al., 2013; Neuman et al., 2013). However,
rather than focus simply on cross-sectional variation in cash effective tax rates (Cash ETR), we
narrow in on the most extreme negative portion of the volatility distribution. Defining downside
tax risk as the probability of a loss event is consistent with other concepts of risk in accounting
and finance, such as extreme downside risk (e.g., Huang et al., 2012), crash risk (e.g., Jin and
Myers, 2006; Hutton et al., 2009; Kim et al., 2011) and operational risk (e.g., Chernobai et al,
2011; Brown et al., 2012). We believe this design choice provides a more powerful setting to
examine the economic forces that determine tax risk because it allows us to identify actual
negative tax outcomes. Second, we identify firms in the tail of our Cash ETR distribution with
2
disclosed, unfavorable tax settlements as “event” observations. We investigate whether ex-ante
characteristics associated with tax risk can predict these tax loss events using a logistic model.
Third, we examine the abnormal stock market returns associated with our proxy for downside tax
risk, as well as a proxy for GAAP ETR management. We investigate abnormal returns before,
during, and after the year of our tax loss events to determine the extent to which the market
anticipates and/or reacts to the disclosure of tax risk, and whether these reactions differ for firms
that do or do not appear to manage the financial reporting of the tax risk through tax expense.
In order to identify our potential tax loss event sample, we measure whether a firm’s
observed annual Cash ETR value exceeds its ten-year firm-specific mean by more than two
standard deviations for the industry (i.e., 4-digit NAICS classification), yielding 284 firm-year
observations over a period from 2002 to 2012. We recognize that an extreme Cash ETR is a
noisy proxy because this ratio may vary for many reasons, including mismatched numerator and
denominator. To address this concern, we explore the annual reports for observations above our
firm-specific Cash ETR value threshold to determine whether available disclosures support the
presence of a loss (i.e., disclosure of a settlement with tax authorities). We identify an event
sample of 52 observations for firms that unfavorably settle with tax authorities.
We estimate a logistic regression on a parsimonious set of tax risk factors to predict the
likelihood of a tax loss event in both a broad sample and a one-to-five matched sample (based on
fiscal year and four-digit industry classification). We document that, as expected, this likelihood
is increasing in higher Cash ETR volatility, the presence of foreign (i.e., multinational)
operations, mergers and acquisitions activities, the number of business segments, and the
existence of SOX 404 internal control weaknesses, and the likelihood is decreasing in the
magnitude of R&D and book-tax differences. The negative relation between book-tax differences
3
and the probability of a loss event suggests that, on average, firms begin to accrue the loss in the
year prior to its payment.
Overall, we find that the probability of an unfavorable tax settlement is indeed greater for
our event sample and that the factors in our multivariate model predict this probability with
reasonable accuracy. To provide further clarity regarding the predictive ability of our model,
given the rarity of tax loss events across our full sample, we construct and analyze a “downside
tax risk” DTR-score. Similar to Dechow et al. (2011), we scale the predicted probability of a tax
loss event by the unconditional probability of that event to create our DTR-score. DTR-scores
above 1 reflect higher downside tax risk. We find that, for our full sample, 63% of our event
observations are in the highest quintile of DTR-score and that we correctly classify 72% of our
event and non-event firm-year observations as either above 1 or below 1, respectively.
Our final set of analyses is designed to test our hypotheses that the market does not price
the tax loss event firms negatively prior to the event year and that in the event year, the market
reacts most negatively for firms that do not manage that risk well. To do so, we examine the
abnormal returns of our event and control firms in a four factor Fama-French (1993) model
around the years our tax loss events are disclosed. We rely on a 2x2 matrix of firms – those in
our tax loss event sample or our control sample interacted with those above or below the median
GAAP ETR Volatility. ETR Volatility is measured as the ten-year firm average standard deviation
of ETR less the ten-year industry average standard deviation of ETR. It reflects the degree to
which firms manage the financial reporting of tax risk. Our evidence supports the hypothesis that
the market anticipates the impending tax loss events, but the negative market reaction continues
in the year of the extreme negative tax event. We fail to find a difference across firms with
differing ETR Volatility. Ongoing work will continue to improve the specification of these tests.
4
Our study complements Gleason and Mills (2002), who focus on the likelihood that a
firm discloses a contingent tax liability. We move beyond the contingency to predict which firms
with high cash effective tax rates actually settle contingent liabilities, and whether and how the
market reacts. Furthermore, our prediction of tax loss events does not necessitate that a firm pre-
disclose or accrue any contingent tax liability; it is sufficient for the firm to acknowledge an
unfavorable settlement with tax authorities in the year of settlement only. To the extent any lack
of disclosure or accrual is considered poor tax risk management, our subsequent market tests
speak to that argument. Put simply, our model predicts significant, unfavorable tax contingencies
– recognized or unrecognized, disclosed or undisclosed – that come to bear on the firm.
More broadly, our study sheds light on the usefulness of financial statement disclosures
for assessing downside tax risk (DTR). We supplement our statistical measure of a tax loss event
with information about the outcome found in annual reports. Our loss measure is accurately
predicted by publicly available, prior year information. The results from our prediction model
suggest that annual reports do provide information regarding tax strategies in general and
regarding tax risk more specifically. However, although our model is a useful predictor of ex-
ante tax risk, results from our market tests suggest that investors typically do not react fully to
these cues prior to the disclosure of a tax settlement.
Finally, this research contributes to the relatively recent literature that examines the tax
risk of the firm. While the practice community has been concerned with, and encouraged clients
in, managing tax risk (see for example, Herskovitz, 1997), little empirical research predates
Guenther et al. (2013) and Neuman et al. (2013). We add to this literature by exploring the
extreme downside tail of tax risk. By focusing on the significant negative risk ex-post, we are
able to develop a prediction model that can be readily applied by market participants and in
5
future research. Furthermore, the variables that aid in the prediction of downside tax risk provide
new insight into the factors that create risk for the firm. This research complements concurrent
research that tries to understand this underexplored element of firm risk.
The remainder of the paper is organized as follows. Section 2 provides background for
our focus on tax loss events. Section 3 provides details of our research design, including our
identification of tax volatility and tax settlements, our prediction model and our abnormal returns
model. Section 4 reports the univariate and multivariate analyses of our prediction model.
Section 5 reports tests of the market reaction to tax loss events and our downside tax risk proxy.
Finally, Section 6 concludes with a summary of the paper.
2. Background
2.1 Large Tax Payments as a Proxy for Downside Tax Risk
Managing corporate tax exposures has become increasingly important within firms’
governance practices. Recent regulation, legislation and standard-setting – such as the Sarbanes-
Oxley Act and FIN 48/ASC 740 – have responded to greater public desire for scrutiny of
corporations, corporate managers and directors with respect to their tax choices. Surveys of tax
executives and managers show an increasing percentage of firms are implementing tax risk
management practices to facilitate informed decision-making (Wunder, 2009; Lavermicocca,
2011). Corporate decision-makers have heightened concern about the risks involved in their tax
strategies. Furthermore, Wunder (2009) documents that U.S. and foreign-owned multinational
corporations report higher aversion to tax risk as a result of the increased focus on tax risk
management.
While there seems to be increased attention, how do external stakeholders view risk,
generally, and how do they conceptualize tax risk, specifically? Further, how can one identify
6
varying tax risks across firms? McGuire et al. (2013) consider the mean and variability of tax
payments of their sample; these values suggest that some firms maintain a consistently low level
of taxes paid while others maintain a consistently high level of taxes paid. Whether high or low
tax, firms appear to have consistent tax volatility. Does maintaining low tax volatility ex post
equate to low tax risk? Guenther et al. (2013) also examine tax rate variability and compute the
standard deviation of cash effective tax rates (Cash ETR). They show that this measure of risk is
related to the firms’ overall stock price volatility, and that the level of Cash ETR and the level of
book tax differences are not related to stock price volatility. Both Guenther et al. (2013) and
McGuire et al. (2013) advance our understanding of tax risk by incorporating the second moment
of the Cash ETR distribution.
Empirical work on asset pricing has reached beyond the assumption of a normal
distribution for stock returns. Dittmar (2002) and Moreno and Rodriguez (2009) incorporated
third and fourth moments into asset pricing models and showed that these were important to
understanding pricing as well. Building on the notion that significant declines are not viewed by
market participants symmetrically to significant gains, a Value at Risk (VaR) metric was
developed to specifically capture the lower tail risk. Bali et al. (2009) finds that such a measure
of risk is positively related to expected returns. Huang et al. (2012) extends this literature by
developing an empirical proxy for the extreme downside risk (EDR) that captures both the
skewness and the fat tails. They show that both the VaR measure and their EDR measure are
idiosyncratic risks that are positively and mutually related to expected returns.
Thus, recent empirical efforts suggest that markets price the risk of unusual negative
events, in spite of opportunities to diversify many firm-specific risks. While there are many
external forces that may lead to a negative event for the firm, its tax positions can expose it to
7
significant, unforeseen tax payments, both as a result of payments for past transactions, and the
required changes that are now necessary in the future. For example, international transfer pricing
cases can result in very significant settlements (e.g., GlaxoSmithKline; Reed, 2006). In addition,
if the transfer prices used in the past are not acceptable, the required change will lead to higher
future taxes as well.
Therefore, we consider tax risk from this perspective. We explore tax risk as the
probability of a loss event—that is, an identifiable event in which a negative tax outcome occurs.
This conceptualization of tax risk is akin to extreme downside market risk (Ang et al., 2006;
Mitton and Vorkink, 2007; Bali et al., 2009; Boyer et al., 2010; Huang et al., 2012). These
studies emphasize that the negative side of the distribution (e.g., the negative stock returns in the
extreme downside risk literature) are more common than expected based on traditional assumed
distributions. In particular, the distributions are skewed and negative tails are fatter than
expected, and that such risk is priced by market participants. The focus of study on so-called tail
events is also evident in the literatures on crash risk (e.g., Jin and Myers, 2006; Hutton et al.,
2009; Kim et al., 2011) and certain forms of operational risk (e.g., Chernobai et al, 2011; Brown
et al., 2012). Each of these areas of inquiry are consistent with anecdotal evidence that managers
are most concerned about large negative outcomes (e.g., March and Shapira, 1987).
In our setting, management establishes their tax planning strategy, which is private
information. Through financial reporting choices around their tax expenses, these managers can
mask the exposure to large tax payouts they undertake, or may simply be unaware of them.
However, risky strategies can eventually yield a significant, unfavorable tax payment. To the
extent that the payment is significant, the amount of taxes paid in the year is high relative to a
8
joint firm-industry average (e.g., n standard deviations from the mean). This occurrence gives
rise to our proxy for a potential tax loss event.
To confirm that the extreme outcomes are indeed loss events, we consider whether firms’
disclosures are relatedly informative. If management holds news until a negative tax settlement is
likely and additional tax payments are made, then disclosures in financial reports should include
discussion of the tax event. Investors can use this information, regardless of its timing, to adjust
their value of the firm. On the other hand, Gleason and Mills (2002) find that U.S. firms
infrequently disclose tax contingencies. Thus if management views the cost of disclosure too
high, perhaps due to potential future dealings with tax authorities, then the related disclosure
could be purposefully less informative, if informative at all, about the settlement. Alternatively,
the event may not be a tax loss at all. Considering that the annual value of taxes paid and pre-tax
book income are inherently volatile (Dyreng et al., 2008), an increase in taxes paid during a
given year could be due to a number of non-events, including taxes on a substantial divestiture or
simply the timing of installment payments. Ultimately, if investors recognize differences in the
information content of tax disclosures that accompany unusually large tax payments, they should
assess the underlying tax risk differently. Furthermore, if factors that imply a high probability of
a tax loss event are evident ex-ante, it is possible to use that information to predict tax risk even
in the absence of a disclosed loss event.
The emerging literature, however, calculates tax risk as a symmetric construct. Measures
of volatility (or tax aggressiveness) are evaluated at the firm-year level using cross-sectional tests
(e.g., McCarty 2012; Guenther et al., 2013; Hutchens and Rego, 2013; Neuman et al., 2013), and
often with firm fixed effects. In our research design, we consider implications based on an
asymmetric form of tax risk: downside tax risk. How the market reacts to this risk of negative
9
outcomes is unclear. If the risk is known, it will require a greater return, even though it is
idiosyncratic (consistent with the VaR and EDR literature). Alternatively, if the risk is largely
hidden by management, then its revelation will yield a negative return as the additional risk is
incorporated into the firm’s value. We predict that there is a relation between the market returns
and downside tax risk.
H1: Abnormal stock returns are related to the realization of tax loss events.
2.2 Downside Tax Risk and Tax Risk Management
In this section, we consider how firms manage downside tax risk, specifically through
their financial reporting risk. According to Neubig and Sangha (2004), the sources of tax risk
include compliance, legislation, technical tax issues, financial reporting, etc. Similar risks are
identified in the tax risk management guide of PricewaterhouseCoopers (2004). Yet despite
documentation that firms are developing specific tax risk management policies (Wunder, 2009),
generally corporate outsiders cannot evaluate the individual components of these policies.
Stakeholders may be able to evaluate effective tax rate management using ETRs that are
readily available in financial statements. Tax professionals interviewed in a study by Mulligan
and Oats (2009) identified tax reserves as a clear example of risk management. Thus managing
an ETR is a form of tax risk management because the ETR is a function of the firm’s tax
reserves, tax accruals and adjustments to valuation allowances. Our predictions of how ETR
management interacts with the probability of a tax loss event follow.
First, consider more generally the relation between hedging and risk documented in the
literature (e.g., Froot et al., 1993; Tufano, 1996; Guay, 1999; Zhang, 2009). These studies
document that firms use derivatives and other financial instruments to hedge business risks and
ultimately maintain or enhance firm value. Similarly, we would expect firms to use tax reserves
10
and accruals to manage the volatility of their ETR and underlying tax risks. Firms that maintain a
consistent level of ETR should be able to avoid unwanted shocks to earnings, whereby smoother
earnings provide value to stakeholders (Grant et al., 2009).1
Nevertheless, we expect there is cross-sectional variation in whether firms with a tax loss
event will effectively anticipate and manage that risk through their ETR. Although Gleason and
Mills (2002) find that both the disclosure of and accrual for tax loss contingencies increase with
the size of the expected loss, such actions are observable for only the largest claims. They find
that many claims that would be material under SFAS No. 5 contingent loss reporting rules are
infrequently disclosed, potentially leading investors to assume the absence of significant
contingencies. Thus, there are two possible predictions. One, shareholders are unaware of the
downside risk that they bear until the tax loss event (and settlement amount) is disclosed by the
firm. Two, shareholders are aware of the risk, but the disclosure resolves the uncertainty and the
price reacts to the change in probability of the loss (to 1, in this case). Alternatively, to the
extent that some firms effectively manage this risk through their ETRs, we expect that the market
will value firms with a tax loss event and weak tax risk management (of the ETR) negatively
relative to firms with a tax loss event and strong tax risk management. If the market participants
are simply updating their probability of the loss, then tax risk management would not
differentially affect returns. More specifically, we state the following hypotheses in null form:
H2: Firms with a tax loss event and volatile firm-level ETR have the same abnormal stock returns as firms with a tax loss event and nonvolatile firm-level ETR in the event period, compared to non-event firms.
1 We view the volatility of ETR as tax risk management, rather than earnings management (for a review of the latter, see Hanlon and Heitzman, 2010). Like McGuire et al. (2013), who compare concepts of sustainable tax avoidance and tax minimization, we are most interested in a consistent strategy that persists over time rather than a decision that affects primarily a given period. ETR volatility (over a ten-year period) is a proxy for this consistent strategy.
11
3. Research Design
3.1 Identifying Downside Tax Risk as a Loss Event
In order to identify firms with downside tax risk and thus the potential for a tax loss
event, we begin by assessing extreme values of Cash ETR, similar to Guenther et al. (2013).2
First, we require that the firm has ten consecutive years of ROA percentages that are greater than
or equal to ½ percent using rolling estimation windows from t-9 to t. This restriction helps to
ensure that we have an economically meaningful sample of annual Cash ETR observations and to
mitigate issues related to small denominators or firms with losses. Second, we calculate the firm-
specific mean and standard deviation of annual Cash ETR for the ten-year window. Ten-year
estimates are consistent with the long-run approach to tax planning. We require ten consecutive
years of data in order to retain these statistics for year t of the rolling window. Third, we use the
firm-specific standard deviation to calculate the rolling ten-year industry average standard
deviation (based on four-digit NAICS industry codes) for each industry and fiscal year of our
sample. Fourth, we calculate our measure of extreme cash payments, the Cash ETR threshold,
for each firm-year as an observed annual Cash ETR value that exceeds the ten-year firm-specific
mean by more than two standard deviations for the industry. Thus any firm-year with a Cash
ETR significantly greater than its mean (e.g., exceeding two industry standard deviations) is
considered a potential tax loss event observation.
3.2 Overall and Tax Loss Event Samples
Our sample selection process begins with all public firms available in Compustat during
the period 2002 to 2012. As reported in Table 1, this initial step yields 99,446 firm-year
observations across 13,982 firms. Next, we require that annual Cash ETR is measureable for
2 We calculate Cash ETR annually as taxes paid divided by pre-tax income less special items (Dyreng et al., 2008). If the denominator is less than zero, we set the observation to missing. We also truncate the distribution between zero and one to reduce the influence of undue outliers that are more likely to represent noise than a tax loss event.
12
each firm-year observation, resulting in a reduction of sample to 47,356 firm-year observations
across 8,644 firms. As we require that firms have an ROA percentage greater than or equal to ½
percent, our sample is further reduced to 41,793 firm-year observations across 8,264 firms after
imposing this restriction. Finally, we ensure that each firm-year observation has ten years of
consecutive data to construct our volatility of Cash ETR measure. This restriction yields a
sample of 8,572 firm-year observations across 1,705 firms that we use to compute our Cash ETR
threshold. Of these firm-year observations, 321 (or 3.7% of the sample) fall in the extreme high
tail and thus are considered potential tax loss events. Consistent with asymmetric risk, we note
conversely that only 132 firm-year observations (or 1.5% of the sample) fall in the extreme low
tail of our Cash ETR threshold distribution.
For each potential tax loss event, we further examine the annual reports of the firm within
the estimation window to determine if a tax settlement occurs or not.3 If the firm has suffered a
loss event we expect to find disclosure about settlements with tax authorities or other relevant
information; these observations represent our tax events. If the firm has not suffered a loss event
we expect to find adequate disclosure or information to validate the high tax payments (e.g., a
substantial divestiture, catch-up for a previous year’s under-installment, etc.). It is possible also
that a firm is not forthcoming or is vague about the reasons for the high tax payments; we
classify these vague observations as a sub-group of the non-event observations. Firms are also
permitted to have multiple tax loss events identified. Our annual report validation reveals 55 tax
loss events. After a final restriction of our sample for observations missing control variable data
we report, in Table 1, 7,438 firm-year observations across 1,507 firms. Of those firm-year
3 We also examine news releases of the firm during the estimation window through Lexis-Nexis. Generally, news releases do not provide identification of settlements with tax authorities that are not disclosed within annual reports.
13
observations, 284 are above our extreme Cash ETR threshold and 52 represent tax loss events for
use in our subsequent prediction model.
Finally, note that our tax loss event identification does not rely on information from
unrecognized tax benefit (UTB) settlements, which are required disclosures under FIN 48 (ASC
740) for fiscal years beginning after December 15, 2006 (and available in Compustat). As we
discuss in greater detail in the Appendix, we do not rely on UTB settlements to identify tax loss
events for two reasons. First, our chosen procedure is applicable in both the pre and post FIN 48
periods, thus using UTB settlements would result in a significant loss of data and
generalizability. Second, the settlement amount reported in a firm’s UTB reconciliation
represents a reduction of a tax liability, yet is does not necessarily represent a current outlay of
cash to tax authorities or an unfavorable settlement. If a firm has remitted tax payments (to
mitigate interest or penalties) in advance of a subsequently favorable settlement then the UTB
settlement disclosure could reflect simply the recovery of that deposit. Such is the case in the
example of International Speedway Corp. in our Appendix. Thus the UTB settlement amount on
its own does not reflect whether a settlement is favorable or unfavorable, or tied to the extreme
Cash ETR value, and relying on it without further, manual validation would introduce noise into
our proxy of downside tax risk.
3.3 Factors that Predict a Tax Loss Event
The probability that a firm incurs a tax loss event – an unfavorable settlement with tax
authorities – is a function of the firm’s tax risk. Recent practitioner literature and academic
literature suggest that tax risk is a function of numerous categories of risk and/or individual
14
factors of risk.4 We develop a parsimonious set of variables using a combination of step-wise
analysis that considers univariate statistics, the fit of the various model iterations and inter-
correlations of variables, and researcher judgment (see Altman, 1968).5 By ensuring parsimony,
our model is easily replicable. All tax risk factors are measured one year prior to the tax loss
event year.
Variables included in the final estimation include the variance of the firm’s annual Cash
ETR, computed over a five-year period. Higher variance in this measure implies more
uncertainty in tax outcomes. We expect a positive relation between Cash ETR VARIANCE and
the likelihood of a tax loss event.
We also measure a firm’s return on assets (ROA) to control for operating performance.
On the one hand, firms with more profitable operations could be expected to undertake tax
planning with a higher level of uncertainty and thus potentially higher payoffs. Yet profitable
firms that manage their tax planning well would not be considered risky. Relatedly, Hutton et al.
(2009) and Kim et al. (2011) show that profitability is negatively related to crash risk, indicating
that loss events in general are more likely for less profitable firms. Therefore, we expect a
negative relation between ROA and the likelihood of a tax loss event.
We include an indicator of foreign operations (FOROPS) in our model for two reasons.
One, the presence of foreign operations suggests a firm is multinational, and we expect
multinational firms to have more tax planning opportunities – and opportunities for risk – than
domestic firms. Two, Wilson (2009) and Lisowsky (2010) find that foreign income is positively
4 For example, see PricewaterhouseCoopers (2004) and OECD (2009) for practitioner literature. See Wunder (2009), McCarty (2012), Guenther et al. (2013), Hutchens and Rego (2013) and Neuman et al. (2013) for academic literature. 5 In addition to other excluded factors discussed in this section, we exclude proxies for tax cushions from our model. Inclusion of a tax cushion estimate, based on the IRS estimate in Gleason and Mills (2002), can explain variation between our event and non-event samples, but drastically reduces our sample because of data limitations.
15
related to the use of tax shelters.6 The inherent likelihood of tax shelters increases downside tax
risk because tax shelters are targeted by tax authorities. For these reasons, we expect FOROPS to
be positively related to the likelihood of a tax loss event.
Firm size (SIZE), the log of total assets, is included as a proxy for the potential scale of
tax planning activities. Including SIZE is consistent also with the contingent tax disclosure model
of Gleason and Mills (2002). However, like Gleason and Mills (2002), our sample firms are
inherently large (we require ten consecutive years of data to determine our high Cash ETR
threshold) and thus we may lack sufficient variation on this factor in our multivariate model.
The ratio of research and development expense to total lagged assets (RND) is in our
prediction model because tax legislation provides favorable credits to research and development
expenditures, but tax authorities heavily scrutinize these claims. Furthermore, as noted in Wilson
(2009), firms can create intangible assets through their research and development activities and
Hanlon et al. (2007) find a positive association between intangible assets and tax non-compliance
(i.e., tax deficiencies). For these reasons, we expect RND to be positively associated with the
likelihood of a tax loss event and thus tax settlements.
We include measures for merger and acquisition activity (MNA) as well as the number of
business segments (NSEG) in our model. Business combinations reflect tax risk because they can
increase the complexity of operations, the tax planning scope and opportunities of the firm, and
the compliance risk surrounding tax deductions and benefits tied to the business combination.
Similarly, a higher number of business segments can increase complexity and opportunities also,
6 Although we measure foreign income (Compustat variable PIFO), including it and a foreign operations indicator (1 if foreign income is positive, 0 otherwise) in our model significantly increases multicollinearity. Thus we include only the foreign operations indicator. Note also that controlling for tax havens does not improve the fit of our model, nor are tax havens a significant predictor, likely due to the amount of missing observations tied to this variable.
16
thus increasing the potential downside tax risk of the firm. Therefore we expect both MNA and
NSEG to be positively associated with the likelihood of a tax loss event.
We control for internal control weaknesses also. Bauer (2013) documents that certain
ICW disclosures reflect weak risk management practices and thus we expect ICW to be
positively associated with the likelihood of a tax loss event. Although both ICWs and tax-related
ICWs are significant predictors of a tax loss event in univariate analysis, we include the more
general ICW proxy in our model. As tax-related ICWs are more infrequent than ICWs in our
sample, they can predict a tax loss event perfectly after matching of control firms on industry,
and would be dropped subsequently from our prediction model. We utilize matched-control
samples to help examine our DTR-score and the abnormal returns of tax loss event firms.
Finally, we include a measure of book-tax differences (BTD) in our model. Per Mills
(1998), proposed IRS audit adjustments increase as BTDs increase. Thus BTDs can act as a
signal to the IRS to investigate firm tax positions, and so we would expect BTD to be positively
associated with the likelihood of a tax loss event.
3.4 Prediction Model and DTR-score
Our primary analysis of asymmetric or downside tax risk predicts the likelihood of a tax
loss event. The output from our prediction model is then adjusted to a scaled probability that acts
as a proxy for downside tax risk, creating a “downside tax risk” score, or DTR-score.
Conceptually, our prediction method is inspired by Altman’s study (1968) that predicts
bankruptcy (i.e., discriminant analysis). Empirically, however, our method more closely follows
the tax liability prediction model of Gleason and Mills (2002) and tax shelter prediction models
in studies by Wilson (2009) and Lisowsky (2010) (i.e., logistic regression models). The
17
construction of a DTR-score is consistent with research by Dechow et al. (2011) on the
likelihood of fraud and earnings misstatements.
Furthermore, our empirical model reflects a discrete hazard model; essentially, we
analyze the time until a tax loss event (i.e. a failure) occurs. Consistent with Beck et al. (1998),
and as explained below, we control for temporal dependence in our data to complete the
identification of the hazard model and to prevent violation of the independence assumptions of a
standard logistic model. This concern arises because our sample of firm-year observations allows
for variation in both the time-series and cross-section.
We predict the probability of a tax loss event in year t using firm i characteristics
measured in the prior year. For ease of illustration, our following logistic hazard prediction
model does not include firm or time subscripts:
ln PTaxEvent1- PTaxEvent
= α+βX+ ε (1)
where PTaxEvent =1
1 + 𝑒−(α+βX+ ε) = the probability the firm will have an unfavorable tax settlement
with the tax authorities; and
βX = β1CASH_ETR_VARIANCE + β2ROA + β3FOROPS + β4SIZE + β5RND + β6MNA +
β7NSEG + β8ICW + β9BTD + β10DURATION + β11DURATION2 (2).
See Table 2 for full details of our variable definitions. The variables DURATION and
DURATION2 are included to control for temporal dependence, where DURATION reflects the
time trend of our sample period (i.e. DURATION equals 1 for 2002 and equals 11 for 2012). The
quadratic specification uses fewer degrees of freedom than the inclusion of dummy variables and
is thus also more parsimonious and more easily replicable. We also adjust the standard errors by
clustering by firm.
18
In order to construct our DTR-score, we scale PTaxEvent by the unconditional probability of
a tax loss event occurring in our sample period. As described previously, our event sample
comprises 52 observations determined by examining annual reports for disclosures about tax
settlements, and we estimate our prediction model on a full sample of 7,438 firm-year
observations. Thus the unconditional probability of a tax loss event for our full sample is 0.70%.
Although our tax loss events are rare, the use of prediction models and/or hazard models in such
contexts is not uncommon. For example, Shumway (2001) uses a discrete logistic hazard model
to improve on Altman’s (1968) prediction of bankruptcy events. Also, Dechow et al. (2011)
develop a logistic model to predict the likelihood of financial misstatement; the unconditional
probability of misstatement in their sample is 0.37%.
Our full sample tests demonstrate the broad applicability of our model. However, to
facilitate our analysis of the market reaction to tax loss event firms, and to perform subsequent
out-of-sample robustness analysis, we also estimate our prediction model and DTR-score based
on a matched control sample. To identify this control sample, we utilize fiscal year and 4-digit
NAICS industry to determine a maximum one-to-five match for our event sample observations.
3.5 Market Reaction – Abnormal Returns
To test Hypotheses 1 and 2, and examine the market reaction of stakeholders to tax loss
events, we perform tests of monthly abnormal returns in a four-factor Fama-French (1993)
model. We assess the market reaction to our tax loss event and control firms conditional on the
strength of ETR risk management, which we proxy for using a firm-level, industry-adjusted
measure of ETR volatility. We perform separate regressions for the years preceding, including,
following and surrounding our event year. Our empirical model, based on Wilson (2009), is as
follows:
19
XRit = β1TaxEvent*WeakRMi + β2Control*WeakRMi + β3Control*StrongRMi +
β4Tax_Event*StrongRMi + bMKTRFt + sSMBt + hHMLt + wUMDt + εit (3).
See Table 2 for full details on our variable descriptions. XR and MKTRF represent
abnormal monthly firm and market returns, respectively, and SMB, HML and UMD represent
size, book-to-market and momentum premia, respectively. We interact a dichotomous measure
of tax loss events (TaxEvent for event firms, Control for selected control firms) with a
dichotomous measure of ETR risk management (StrongRM and WeakRM) and include all four
two-way interaction variables in the model (and omit the intercept).
Our ETR risk management proxy is based on our measure of ETR Volatility. ETR
Volatility is measured as the standard deviation of a firm’s annual ETR less the four-digit,
NAICS industry standard deviation of annual ETR, where standard deviation is constructed over
a ten year period. Firms with lower volatility are considered stronger risk managers. Thus we
label firms with below-median ETR Volatility as StrongRM and those with above-median ETR
Volatility as WeakRM. We then interact these conditions with our tax loss event conditions to
construct our four two-way interaction variables of interest.
4. Results for Tests of the Prediction Model
4.1 Descriptive Statistics – Sample Distributions
As a first set of analyses, we examine the distribution of observations by our firm-level
Cash ETR threshold. We compare our overall control sample of observations (7,154 firm-years,
with four-digit industry) to both our aggregate extreme Cash ETR sample (i.e., 284 potential tax
loss event firm-years) and our validated extreme Cash ETR sample (i.e., 52 event firm-years, 192
non-event firm-years and 40 vague, non-event firm-years). Table 3, Panels A through D
20
respectively, shows distributions across fiscal year, 2-digit industry (rather than 4-digit, for
brevity), deciles of size and deciles of profitability.
Table 3, Panel A, shows that all years from 2002 through 2012 have a potential tax loss
event. However, most likely due to the economic downturn, a disproportionately larger number
of potential tax loss events occur in the middle period of our sample years.7 Our tax event sample
rises through 2007 and then declines again. Table 3, Panel B, shows that every industry has a
firm with a tax loss event; we retain industries with at least one extreme Cash ETR observation.
Also, relatively high proportions of large observations occur in agriculture (code 11; 10.5% of
the total sample), construction (code 23; 9.4% of the total sample) and extractive (code 21; 8.4%
of the total sample) industries. Our tax event sample also has a high number of events from the
agriculture industry (4.8% of the total sample).
Table 3, Panel C, gives some indication that the extreme Cash ETR values are found
among firms that are smaller. The lowest three deciles of size have over 35 observations each.
However, the distribution of our tax event sample is much more even across the size deciles; in
fact, the highest number of tax event observations (i.e., 7) occurs in the smallest and largest
deciles. Thus, size does not appear as significant a factor of tax loss events as of high Cash
ETR.8 On the other hand, profitability does appear to be an important factor to tax loss events.
Table 3, Panel D, shows that 131 of the 284 potential tax loss events (i.e., 46%) are reported in
the lowest two deciles of profitability. Similarly, 26 of the 52 tax events (i.e., 50%) are reported
7 The economic downturn would affect the Cash ETR values by decreasing the denominator, making the observed ratio higher. While we verify each observation, the larger initial sample due to this statistical phenomenon will yield a larger number of real events. 8 A chi-square test fails to reject the null hypothesis of no association between size decile and loss event classifications at the 10% level of statistical significance.
21
in the lowest two deciles of profitability also. As described in Sections 3.2 and 3.3, we control
for firm size (i.e., SIZE) and profitability (i.e., ROA) in our prediction model.
4.2 Descriptive Statistics – Univariate Analysis
Building on the sample distribution analysis, we provide traditional descriptive statistics
for each independent variable in our model in Table 4. These statistics are provided for variables
measured in the year prior to the event period. We include both the full sample of non-tax event
observations, as well as the matched sample, as described above. Finally, we report the p-value
from a t-test of means for each variable.
The variance of cash ETR, the presence of ICWs and book-tax differences have
differences in means at the 1% level of statistical significance. In comparison to the full sample,
the number of segments is also different at the 5% level of statistical significance, but is very
similar in comparison to the matched sample. On the other hand, the mean return on assets and
mean R&D are different in comparison to the matched sample at the 5% level, whereas these
means are different from the respective means of the full sample at only the 10% level of
statistical significance. The differences in means are generally in the direction we predicted,
except RND and BTD are lower for the tax events, contrary to our prediction of a positive
relation.9
4.3 Logistic Regression Results
In this section we report the results from our primary logistic regression models that
predict a tax loss event. Table 5 contains two columns of results based on which sample is used
as the control: the first column presents results based on the full sample of observations; the
second column presents results using the matched sample as a control. Within each column, by
9 In untabulated analysis, no significant difference exists, statistically or in magnitude, between event firms and control firms on the basis of intangible assets. Intangible assets are frequently linked to R&D activities.
22
variable, the first value reported represents the coefficient; the second value (in parentheses)
represents the z-statistic, and; the third value (in straight brackets) represents the coefficient
interpreted as a marginal effect.
The first column of results in Table 5 demonstrates that our logistic model provides a
reasonable prediction of the likelihood that a firm has a tax loss event (i.e., a settlement with tax
authorities). The pseudo R2 value of 0.12 and chi2 statistic of 107.4 imply that the model is well
fit. Furthermore, the area under the Receiver Operator Characteristic (ROC) Curve is 0.80,
further validating the fit of the model. This ROC Curve statistic implies that for a random pair of
firms – one with a tax loss event and one without– our model would correctly identify the firm
with a tax loss event 80% of the time (see Hosmer and Lemeshow, 2000). In contrast, a ROC
Curve statistic of 0.50 would indicate that our model has no greater predictive power than chance
(i.e., 50%).
In the first column of Table 5, using a 5% cut off on two-tailed tests of statistical
significance, seven of the nine firm characteristics are statistically significant; ROA and SIZE are
not statistically significant. Cash ETR VARIANCE, FOROPS, MNA, NSEG, ICW, and
DURATION are positively related to the likelihood of a tax loss event, as expected. Thus firms
with greater multinational scope and firms undergoing a business combination, both of which
can imply complexity and more uncertain tax opportunities, are more likely to have an
unfavorable settlement with tax authorities.
The other three statistically significant coefficients in the first column of Table 5 are
negatively related with our dependent variable. DURATION_SQ is negative, as predicted,
consistent with the effect of DURATION declining as the duration grows. Consistent with
univariate results, firms with lower R&D expense are more likely to have a tax loss event. It is
23
possible that these firms, on average, do not have the tax shield created by R&D and so take on
riskier tax activities. BTD is also estimated to have a negative relation with the tax event
indicator. Although this result is unexpected, evidence in Table 3 Panel D documents that a high
proportion of firms with tax loss events have low profitability. If profitability is low, then book
income may be low relative to taxable income, particularly if tax expense (or ETR) is higher
pursuant to an accrual for settlement. Untabulated analysis documents that 46% of the tax loss
event firm-years in the bottom quartile of the range of BTD (i.e. BTD < 0.004) have below-
median levels of profitability.
Looking at the second column of Table 5, the pseudo R2 value indicates that the
explanatory power of the model improves with the matched sample. The area under the ROC
Curve increases slightly, to 82%, also. The individual coefficients differ across the two
regressions, but are generally similar in sign and statistical significance. However, MNA and the
two DURATION coefficients are no longer statistically significant at the 10% level using a two-
tailed test.
Broadly speaking, the evidence in Table 5 demonstrates that the firm characteristics we
have chosen provide reasonable predictions of the likelihood that a firm experiences a significant
tax payment in the year following. These firms have encountered an unfavorable tax outcome,
evidence of the downside risk inherent in their tax planning decisions.
4.4 Examining the Estimated Predictions In-Sample: DTR-score
As we view the existence of a tax event to be an extreme negative event, consistent with
higher probability of such outcomes, we use the prediction model in the previous section to
potentially identify other firm-years in which such risk existed, but the extreme tax payment is
not observed. To achieve this objective, we follow Dechow et al. (2011) and compute the
24
predicted probability from our fitted values for our full and matched-sample model. These
predicted values are then deflated by the unconditional probability of the event to yield a ratio,
our DTR-score. The ratio is greater than 1 if the predicted probability exceeds the unconditional
predicted probability.
Table 6 tabulates these scores by quintile for both the full sample and the matched
sample. Each quintile is segmented into the observed tax events and the observations without an
observed tax event. The table reveals that the observed tax events are increasing across the
quintiles, with 33 and 26 events (63% and 50% of the events) found in the highest quintile when
using the full sample and matched control sample, respectively. The sensitivity, that is, the
observed events correctly identified using a cut-off score of 1, equals 73% and 75%,
respectively, for the two analyses. These sensitivities are similar to those reported by Dechow et
al. (2011). The type I error rates (that is, non-event observations with a score greater than 1) are
28% and 24%, respectively. This compares favorably to the rates in Dechow et al. of 36%. One
would expect a reasonable proportion of Type I errors since many firms that have a high risk of
the extreme tax payment are not caught by tax authorities.
The evidence in Table 6 suggests that the prediction model provides meaningful
separation of firms that have a tax event and firms that are in the control sample. These analyses
support the conclusions drawn from the ROC curve described above. As further corroboration of
our DTR-score, we provide Figures 1 and 2. These figures show the distribution of DTR-score
for event and matched firms in the t-4 to t+4 window around the event (match) year. Figure 1
reports the median DTR-score for each sample, and Figure 2 reports the percentage of DTR-
score values above 1.0.
25
Figure 1 shows that the median DTR-score is higher for event firms relative to matched
firms in the t-1 through t+2 period. Given that DTR-score is calculated from firm characteristics
measured in the prior period, these scores reflect higher median downside tax risk in the t-2
through t+1 period. Thus the downside risk associated with the likelihood of tax loss events
appears to persist over at least four years, and is highest in the period of the event. Similarly,
Figure 2 shows that the percentage of DTR-scores above 1.0 is also near its peak in the event
period, and that the separation between event and match firms persists over a similar four-year
period.
5. Market Reaction Tests
5.1 Abnormal Returns with Tax Event Firms matched to Primary Control Sample
In this and the following section, we report the results of our market reaction tests of
Hypotheses 1 and 2. That is, we look at the abnormal returns from a four factor Fama-French
(1993) model for a sample of event and non-event firms to determine whether the risk of a tax
loss event is valued in the pre-event period, and whether tax loss event firms with high ETR
volatility have lower returns than tax loss event firms with low ETR volatility in the event
period. We also examine the post-event period and aggregate period for completeness.
In columns (1) through (4) of Table 7, Panel A, we report results from the abnormal
returns test where our tax loss event firms are matched with control firms based on year and
industry (as identified in Section 4.3). Column (1), the pre-period, shows significant (1% level,
two-tailed), positive coefficients for Control*WeakRM and Control*StrongRM. This suggests
that for all control firms, the returns are positive, whereas the returns are not positive for the
event firms. An F-test that compares the two coefficients for event firm-years to those of control
26
firm-years show a difference that is statistically different at the 10% level using a two-tailed test.
This provides support for Hypothesis 1 using data from years before the event.
Using data for the year of the event, the coefficient on TaxEvent*WeakRM is negative
whereas the coefficient on Control*WeakRM is zero. The coefficients for the StrongRM
interactions are positive for the Control sample and not statistically significant for the TaxEvent
sample. Like the year prior, the test of Hypothesis 1 rejects the null at the 5% level of statistical
significance in the event period.
In a test of Hypothesis 2, the coefficients of TaxEvent*StrongRM and Control*StrongRM
are significantly different from one another at the 10% level using a two-tailed test. The
difference in coefficients across the WeakRM groups is not statistically significant at
conventional levels. Thus, using data from t=0, the negative response is evident within the strong
risk management firms whereas it is not evident in the weak risk management firms.
Using data from the prior year, the returns for the TaxEvent sample is negative relative to
the Control sample only for the WeakRM group (at the 10% level of statistical significance, one-
tailed); the difference across the StrongRM group is not statistically significant.
Results reported in columns (3) and (4) of Table 7, Panel A, show that in the post-period
and for all years for the firms, there are no differences across the coefficients.
5.2 Abnormal Returns with Predicted Values in the Broader Sample
To assess the ability of our model to predict the out-of-sample likelihood of a tax event
we also re-estimate our abnormal returns model where our tax loss event is defined using our
prediction model. Specifically, we estimate PTaxEvent for all of the firms not matched to our event
sample, using the coefficient values from the second column of Table 5. We label this predicted
27
probability PDTR. Firm-year observations in the highest decile of the distribution of PDTR are also
labeled as predicted TaxEvent observations (i.e. PrTaxEvent, otherwise labeled PrControl).
In column (1) of Table 7, Panel B, we report the re-estimation of Equation (3) after the
interaction of PrTaxEvent and PrControl with StrongRM and WeakRM. In column (2), we adjust
the model to include PDTR, PrTaxEvent and the interaction of these two variables (as well as an
intercept). In column (3), we supplement the column (2) model by adding StrongRM and the
respective two-way interactions and three-way interactions between StrongRM, PDTR and
PrTaxEvent. As no actual event date exists for this sample, we re-estimate the models across the
full sample period (i.e. 2002-2012), rather than in the t-1 to t+1 window around an actual event.
The results from column (1) of Table 7, Panel B provide evidence that across the broad
hold-out sample, observations with a higher predicted downside tax risk generate higher returns
for investors. The returns for the predicted TaxEvent sample are positive relative to the Control
sample for both the WeakRM and StrongRM groups (at the 5% level of statistical significance,
two-tailed). The difference within the predicted TaxEvent group across WeakRM and StrongRM
is not statistically significant, however. Broadly, this evidence is consistent with Hypothesis 1, in
that greater risk requires a greater return.
The results from columns (2) and (3) of Table 7, Panel B are weaker. None of the
coefficients of interest are statistically significant in column (2). In column (3), the return to
investors appears to increase as the probability of downside tax risk increases for firms with low
ETR volatility (i.e. StrongRM*PDTR is positive and statistically significant at the 5% level).
6. Conclusion
This study contributes to the literature on tax risk, yet from a different perspective than
extant research. We consider downside tax risk as the probability of a loss event, in which an
28
unfavorable and perhaps unexpected tax settlement occurs. Like studies of extreme downside
risk (e.g. Huang et al., 2012), crash risk (e.g., Jin and Myers, 2006; Hutton et al., 2009; Kim et
al., 2011) and operational risk (e.g., Chernobai et al, 2011; Brown et al., 2012), we focus on the
observations in the extreme negative tail of the distribution, where the likelihood of a loss is
significantly greater. With this conceptualization in mind, we identify firms with extreme and
high values of Cash ETR (relative to an industry benchmark) and a disclosed tax settlement in
their financial statements as tax loss event firms. Furthermore, we develop a logistic prediction
model from firm characteristics on which we measure a scaled probability (DTR-score) that
reflects the likelihood of a tax loss event and is a signal of downside tax risk. Finally, we
examine the extent to which the market reacts to this downside tax risk.
Empirical results from tests of our prediction model demonstrate that the likelihood of a
tax loss event (i.e., unfavorable settlement with tax authorities) is positively associated with tax
risk factors such as higher Cash ETR volatility, foreign operations, M&A activity, business
segments (i.e. complexity) and SOX internal control weaknesses. This likelihood is also
negatively associated with R&D expenditures and book-tax differences. Using our DTR-score,
we demonstrate that in a broad sample our model correctly classifies the likelihood of a tax loss
event (i.e. DTR-score above 1) for observed events with 73% accuracy. Also, our classification,
Type 1 and Type II error rates compare favorably to other studies that examine the likelihood of
relatively rare events in broad samples (e.g. financial misstatements; Dechow et al., 2011).
Similar classification rates result from a matched sample analysis of our DTR-score, further
validating the strength of our prediction model and our proxy for downside tax risk.
Empirical results from tests of the market reaction to these tax loss events indicate that
the shareholders of these tax loss event firms anticipate the tax loss events, although the negative
29
market reaction appears highest in the year of the extreme negative event. We fail to find a
differential response in the market to firms with high GAAP ETR volatility as compared to low
ETR volatility. In a holdout sample examination of the predicted probability of tax loss events,
we find some evidence that firms with higher downside tax risk generate higher returns for
shareholders. Again, we do not find a differential response across ETR volatility, a proxy for the
strength of tax risk management.
We believe that our study provides valuable insight regarding the nature of tax risk within
firms. In future research, our model could be used not only to predict downside tax risk in a
broader sample but also to help evaluate the extent to which firms effectively manage their tax
risk. More generally, our model and its predicted probabilities should be useful for market
participants and other firm stakeholders in their assessments of firm risk.
30
Appendix Discussion of Tax Loss Event firms and UTB Settlement Data from the post-FIN 48 Period
Our identification of Tax Loss Event firms relies on manual verification of information (i.e. disclosure of unfavorable settlements with tax authorities) contained in 10-K reports. Our sample period spans 2002 to 2012, and thus nearly half of it falls in the post-FIN 48 period – when firms disclosure UTB reserves and settlements in their 10-K reports. Compustat collects and compiles information from UTB disclosures, yet we do not rely on UTB settlements (i.e. TXTUBSETTLE per Compustat) for two reasons. One, observations from the pre-FIN 48 period would be lost, reducing the generalizability and scope of our prediction model. Two, the value of TXTUBSETTLE does not necessarily represent an unfavorable settlement with tax authorities, if it is disclosed or compiled at all. For example, as with International Speedway Corp. below, the value of TXTUBSETTLE could simply represent the financial statement reversal of a previous accrual in which no additional taxes paid are required upon settlement. Thus without additional, manual verification, the UTB settlement data represents a noisy proxy for our categorization of tax loss events, a proxy for asymmetric, downside tax risk.
The following table lists information for firms above our Cash ETR threshold to highlight four different situations that could arise in the post-FIN 48 period, some of which represent tax loss events and some of which represent noise (i.e. non-events). We explain each of these four situations following the table.
GVKEY FYEAR Company Name TXPD TXTUBSETTLE Tax Loss Event 5680 2009 Lowe’s Companies Inc. $1,157M $1M 0 6116 2009 International Speedway Corp. $36.297M $121.532M 0 8530 2010 Pfizer Inc. $11,775M $0M 1
12403 2009 Fred’s Inc. $22.999M $8.6M 1 Lowe’s Companies Inc. Lowe’s reports a UTB settlement but we label it as a non-event observation. Furthermore, this observation is also part of the group with “vague” disclosures. We consider the disclosure to be indicative of a vague non-event because certain information implies a settlement; however it is unclear whether any related amount was paid during the year (and thus contributes to the extreme Cash ETR value observed). The 10-K contains the following excerpt:
“The Company is subject to examination by various foreign and domestic taxing authorities. During 2009, the IRS completed its examination of the Company’s 2004 and 2005 income tax returns, with the exception of certain issues that are presently under appeal.”
International Speedway Corp. International Speedway reports a UTB settlement but we label it as a non-event observation. As indicated in the disclosure below, the majority of this settlement amount represents a return of deposit made by International Speedway in a previous year. Thus the settlement amount that firms disclose represents a reduction of the UTB reserve amount but it does not necessarily represent taxes paid during the year. In this particular case, the settlement is favorable and no additional taxes paid are required. An excerpt from the 10-K is as follows:
31
“The company entered into a definitive settlement agreement with the Internal Revenue Service in connection with the federal income tax examination for the 1999 through 2005 fiscal years on May 28, 2009. As a result of the Settlement, on June 17, 2009, the company received approximately $97.4 million of the $117.9 million in deposits that it had previously made with IRS. In addition, it received approximately $14.6 million in cash for interest earned on the deposited funds.”
Pfizer Inc. Pfizer does not report a UTB settlement but we label it as a tax loss event observation. In its UTB reconciliation (untabulated in this Appendix), Pfizer does not include a line specifically labeled as for “settlements”. Therefore Compustat does not recognize any amount for its TXTUBSETTLE variable. The following excerpt, however, clearly identifies a settlement with tax authorities:
“During the fourth quarter of 2010, we reached a settlement with the U.S. Internal Revenue Service (IRS) related to issues we had appealed with respect to the audits of the Pfizer Inc. tax returns for the years 2002 through 2005, as well as the Pharmacia audit for the year 2003 through the date of merger with Pfizer (April 16, 2003). The IRS concluded its examination of the aforementioned tax years and issued a final Revenue Agent’s Report (RAR). We have agreed with all of the adjustments and computations contained in the RAR.”
Fred’s Inc. Fred’s reports a UTB settlement and based on additional information in its 10-K we classify it as a tax loss event observation. The additional information indicates that taxes paid during the year reflect, in part, amounts paid in settlement. The 10-K disclosure includes the following excerpt:
“…the Internal Revenue Service [finalized] an exam of the Company during 2009 covering fiscal years 2004 through 2007. The Company recorded a reduction to unrecognized tax benefits for settlement of the IRS exam which included tax and interest in the amount of $8.6 million.”
32
References
Altman, E. I., 1968. Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy. Journal of Finance 23(4), 589-609.
Ang, A., Chen, J., Xing, Y., 2006. Downside risk. The Review of Financial Studies 19(4), 1191-1239.
Bali, T. G., Demirtas, K. O., Levy, H., 2009. Is there an intertemporal relation between downside risk and expected returns? Journal of Financial and Quantitative Analysis 44(4), 883-909.
Bauer, A., 2013. Tax Avoidance and the Implications of Weak Internal Controls. Working paper, University of Illinois at Urbana-Champaign.
Beck, N., Katz, J.N., Tucker, R., 1998. Taking Time Seriously in Binary Time-Series-Cross-Section Analysis. American Journal of Political Science 42(4), 1260-1288.
Boyer, C. M., Ciccone, S. J., Swicegood, P., 2010. The search for A dog-free portfolio: Why unions matter to investors. Journal of Business & Economics Research 8(6), 67-71.
Brown, S., Goetzmann, C., Liang, B., Schwarz, C. 2012. Trust and Delegation. Journal of Financial Economics 103, 221-234.
Chen, S., Chen, X., Cheng, Q., Shevlin, T., 2010. Are family firms more tax aggressive than non-family firms? Journal of Financial Economics 95(1), 41–61.
Chernobai, A., Jorion, P., Yu, F., 2011. The Determinants of Operational Risk in U.S. Financial Institutions. Journal of Financial and Quantitative Analysis 46(6), 1683-1725.
Dechow, P.M., Ge, W., Larson, C.R., Sloan, R.G., 2011. Predicting Material Accounting Misstatements. Contemporary Accounting Research 28(1), 17-82.
Dittmar, R. F., 2002. Nonlinear pricing kernels, kurtosis preference, and evidence from the cross section of equity returns. The Journal of Finance 57(1), 369-404.
Dyreng, S., Hanlon, M., Maydew, E., 2008. Long-Run Corporate Tax Avoidance. The Accounting Review 83(1), 61-82.
Dyreng, S., Hanlon, M., Maydew, E., 2010. The Effects of Executives on Corporate Tax Avoidance. The Accounting Review 85(4): 1163-1189.
Fama, E. F., French, K. R., 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33(1), 3-56.
Frank, M., Lynch, L., Rego, S., 2009. Tax Reporting Aggressiveness and Its Relation to Aggressive Financial Reporting. The Accounting Review 84(2), 467-496.
Froot, K. A., Scharfstein, D.S., Stein, J. C., 1993. Risk management: Coordinating corporate investment and financing policies. The Journal of Finance 48(5), 1629–1658.
Gleason, C.A., Mills, L.F., 2002. Materiality and Contingent Tax Liability Reporting. The Accounting Revnew 77(2), 317-342.
Graham, J. R., Raedy, J. S., Shackelford, D. A., 2012. Researching in accounting for income taxes. Journal of Accounting and Economics 53(1-2), 412–434.
Grant, J., Markarian, G., Parbonetti, A., 2009. CEO Risk-Related Incentives and Income Smoothing. Contemporary Accounting Research 26(4), 1029–1065.
Guay,W.,1999.The impact of derivatives on firm risk: an empirical examination of new derivative users. Journal of Accounting and Economics 26, 319–351.
33
Guenther, D. A., Matsunaga, S. R., Williams, B. M., 2013. Tax Avoidance, Tax Aggressiveness and Firm Risk. Working paper, University of Oregon.
Hanlon, M., 2003. What Can We Infer about a Firm’s Taxable Income from Its Financial Statements? National Tax Journal, 56(4), 831-863.
Hanlon, M., Mills, L., Slemrod, J., 2007. An empirical examination of corporate tax noncompliance. In Taxing Corporate Income in the 21st Century, edited by A. Auerbach, J. R. Hines, Jr., and J. Slemrod. Cambridge, U.K.: Cambridge University Press.
Hanlon, M., Heitzman, S., 2010. A Review of Tax Research. Journal of Accounting & Economics 50 (2-3), 127-178.
Herskovitz, D.L., 1997. Managing an IRS audit: Practical insights and techniques. A dialogue. Tax Executive 49, 375-385.
Higgins, D., 2013. Communicating Information Beyond the Tax Footnote: Effective Tax Rate Forecasts in the MD&A. Working paper, University of Connecticut.
Hosmer, D., Lemeshow, S., 2000. Applied Logistic Regression. 2nd edition. New York, NY: John Wiley & Sons, Inc.
Huang, W., Liu, Q., Rhee, S.G., Wu, F., 2012. Extreme downside risk and expected stock returns. Journal of Banking and Finance 36(5), 1492-1502.
Hutchens, M., Rego, S., 2013. Tax Risk and the Cost of Equity Capital. Working paper, Indiana University.
Hutton, A.P., Marcus, A.J., Tehranian, H., 2009. Opaque financial reports, R2, and crash risk. Journal of Financial Economics 94, 67–86.
Jin, L., Myers, S.C., 2006. R2 around the world: new theory and new tests. Journal of Financial Economics 79, 257–292.
Kim, J-B., Li, Y., Zhang, L., 2011. Corporate tax avoidance and stock price crash risk: Firm-level analysis. Journal of Financial Economics 100(3), 639–662.
Lavermicocca, C., 2011. Tax Risk Management Practices and their Impact on Tax Compliance Behaviour – The Views of Tax Executives from Large Australian Companies. eJournal of Tax Research 9(1), 89-115.
Lenter, D., Shackelford, D. A., Slemrod, J., 2003. Public Disclosure of Corporate Tax Return Information: Accounting, Economics, and Legal Perspectives. National Tax Journal, 56(4), 803-830.
Lisowsky, P., 2010. Seeking Shelter: Empirically Modeling Tax Shelters Using Financial Statement Information. The Accounting Review 85(5), 1693-1720.
March, J. G., Shapira, Z., 1987. Managerial perspectives on risk and risk taking. Management Science 33(11), 1404-1418.
McCarty, R.D., 2012. Optimal Tax Risk and Firm Value. Working paper, Northeastern University.
McGuire, S. T., Neuman, S. S., Omer, T. C., 2013. Sustainable Tax Strategies and Earnings Persistence. Working paper, Texas A&M University.
Mills, L.F., 1998. Book-Tax Differences and Internal Revenue Service Adjustments. Journal of Accounting Research 36(2), 343-356.
Mitton, T., Vorkink, K., 2007. Equilibrium underdiversification and the preference for skewness. The Review of Financial Studies 20(4), 1255-1288.
34
Moreno, D., Rodríguez, R., 2009. The value of coskewness in mutual fund performance evaluation. Journal of Banking & Finance, 33(9), 1664-1676.
Mulligan, E., Oats, L., 2009, Tax Risk Management: Evidence from the United States. British Accounting Review 6, 680-703.
Neubig, T., Sangha, B., 2004. Tax risk and strong corporate governance. The Tax Executive 56(2), 114–119.
Neuman, S. S., Omer, T. C., Schmidt, A. P., 2013. Risk and Return: Does Tax Risk Reduce Firms’ Effective Tax Rates? Working paper, Texas A&M University, University of Nebraska-Lincoln, North Carolina State University.
OECD, 2009. Corporate Governance and Tax Risk Management. Forum on Tax Administration, Organisation for Economic Co-operation and Development, Paris.
PricewaterhouseCoopers, 2004. Tax risk management. London: PWC.
Raedy, J. S., Seidman, J. K., Shackelford, D. A., 2011. Is There Information Content in the Tax Footnote? Working paper, University of North Carolina and University of Texas-Austin.
Reed, K., 2006. “GSK Pays ‘Biggest Tax Settlement in U.S. History.” Accountancy Age, September 9, 2006. Retrieved October 8, 2013 from http://www.accountancyage.com/aa/news/1768893/gsk-pays-biggest-tax-settlement-us-history.
Shumway, T., 2001. Forecasting Bankruptcy More Accurately: A Simple Hazard Model. The Journal of Business 74(1), 101-124.
Tufano, P., 1996. Who manages risk? An empirical examination of risk management practices in the gold mining industry. The Journal of Finance 51(4), 1097–137.
Wilson, R., 2009. An Examination of Corporate Tax Shelter Participants. The Accounting Review 84(3), 969-999.
Wunder, H., 2009. Tax risk management and the multinational enterprise. Journal of International Accounting Auditing & Taxation 18(1), 14-28.
Zhang, H., 2009. Effect of derivative accounting rules on corporate risk-management behavior. Journal of Accounting and Economics 47, 244-264.
35
http://www.accountancyage.com/aa/news/1768893/gsk-pays-biggest-tax-settlement-us-historyhttp://www.accountancyage.com/aa/news/1768893/gsk-pays-biggest-tax-settlement-us-history
Figure 1 Median DTR-score
Figure 1 shows the median DTR-score in each year around the t=0 event year for our event and match sample observations. We include all firm-year observations related to both our event firms and our matched control firms in our sub-sample. A maximum of five control firms are matched on fiscal year and four-digit industry classification for each event observation. We estimate our prediction model and scale the predicted probability of a tax loss event by the unconditional probability of an event for the sub-sample to generate the DTR-score for each observation. For each year in the t-4 to t+4 window around the event year or match year, we report the median DTR-score by event firms and match firms.
36
Figure 2 Percentage of DTR-score > 1.0
Figure 2 shows the percentage of observations with a DTR-score above 1.0 in each year around the t=0 event year for our event and match sample observations. We include all firm-year observations related to both our event firms and our matched control firms in our sub-sample. A maximum of five control firms are matched on fiscal year and four-digit industry classification for each event observation. We estimate our prediction model and scale the predicted probability of a tax loss event by the unconditional probability of an event for the sub-sample to generate the DTR-score for each observation. For each year in the t-4 to t+4 window around the event year or match year, we report the percentage of observations with a DTR-score above 1.0 by event firms and match firms.
37
Table 1 Sample Selection Procedure
Criteria Firm-Year
Observations Firm-Level
Observations Public firms in Compustat, 2002-2012 99,446 13,982 Less: observations without an annual Cash ETR value (52,090) (5,338) 47,356 8,644 Less: observations with an annual ROA value less than 1/2% (5,563) (380) 41,793 8,264 Less: observations without ten years of consecutive data to measure Cash ETR benchmarks (33,221) (6,559) Remaining sample to estimate Cash ETR threshold, 2002-2012 8,572 1,705 Less: observations missing data for control variables (1,134) (198) Final sample for prediction model, 2002-2012 7,438 1,507
38
Table 2 Variable Definitions
Panel A – Prediction Model Variables
Dependent Variable Constructs
Criteria
Tax Loss Event = 1 if an observation has an above-threshold value of Cash ETR and financial statements disclose an unfavorable settlement with tax authorities, 0 otherwise. Measured at time t.
Cash ETR threshold = 1 if the difference between the firm-year annual Cash ETR and firm-level ten-year average Cash ETR exceeds 2 standard deviations of the ten-year, four-digit NAICS industry average Cash ETR, 0 otherwise. Measured in rolling windows at time t.
Cash ETR = Per Dyreng et al. (2008), Taxes paid divided by (Pre-tax income less special items); txpd / (pi – spi), per Compustat. Set to missing if the denominator is negative and truncated between 0 and 1. For the ten-year rolling averages of Cash ETR, the sum of the numerator is divided by the sum of the denominator. Annual and rolling average values are measured at time t.
DTR-score = The predicted probability of a tax loss event PTaxEvent (e.g. constructed using the coefficient values from Table 5) scaled by the unconditional probability of a Tax Loss Event in the sample data. Measured at time t, based on data at time t-1.
Independent Variables Cash ETR Variance = The variance of annual Cash ETR over rolling five-year windows per
firm. Measured at time t-1. ROA = Return on Assets; (pi – xi) / lag(at), per Compustat. Restricted to include
values that are greater than or equal to ½% only. Measured at time t-1. FOROPS = 1 if foreign income is positive, 0 otherwise; Foreign income = pifo, per
Compustat. Measured at time t-1. SIZE = Natural log of total assets; Total Assets = at, per Compustat. Measured at
time t-1. RND = R&D expense divided by lagged total assets; xrd / lag(at), per
Compustat. Set to zero if R&D expense is missing. Measured at time t-1. MNA = 1 if observation has mergers and acquisitions (M&A) activity during the
year, 0 otherwise; M&A = aqp, per Compustat. Set to zero if M&A data is missing. Measured at time t-1.
NSEG = Number of segments reported; segments reported = ptis, per Compustat segment file. Set to zero if segment data is missing or firm is not in Compustat segment file. Measured at time t-1.
ICW = 1 if firm-year reports a material internal control weakness (ICW), 0 otherwise; ICW = IC_IS_EFFECTIVE (“N” or “Y”), per Audit Analytics. Measured at time t-1.
BTD = Per Wilson (2009), Book-tax difference is book income less taxable income scaled by lagged assets; [(pi-(txfed+txfo)/stxr)-(tlcf-lag(tlcf))] / lag(at), per Compustat. If txfed is missing, taxable income = (txt-txdi-txs-txo) / stxr, where stxr is the statutory tax rate. Measured at time t-1.
DURATION = A time trend representing the years elapsed from the beginning of the sample period (i.e. 2002 = 1, 2003 = 2, etc.).
DURATION2 = The square of DURATION to complete a quadratic time trend.
39
Table 2 – continued
Panel B – Abnormal Returns Model Variables
Dependent Variable Criteria XR = Firm i’s excess stock return during month t, the stock return minus risk-
free rate [rit – rft]; ret – rf, per CRSP. Independent Variable Constructs
TaxEvent (PrTaxEvent) = 1 if firm i has a Tax Loss Event value of 1 (i.e., above-threshold Cash ETR value and financial statements disclose an unfavorable settlement with tax authorities), 0 otherwise. For holdout sample tests, PrTaxEvent equals 1 if a firm-year observation is in the top decile of DTR-score, 0 otherwise.
Control (PrControl) = 1 if firm i has a Tax Loss Event value of 0 and has been matched to a Tax Loss Event firm (based on event year and four-digit NAICS industry), 0 otherwise. For holdout sample tests, PrControl equals 1 if a firm-year observation is outside the top decile of DTR-score, 0 otherwise.
WeakRM = 1 if firm i has an above-median value of ETR Volatility, 0 otherwise. StrongRM = 1 if firm i has a below-median value of ETR Volatility, 0 otherwise. ETR Volatility = The standard deviation of firm i’s annual ETR less the four-digit, NAICS
industry standard deviation of annual ETR, where standard deviation is constructed over a ten year period. Per Dyreng et al. (2008), ETR equals tax expense divided by (Pre-tax income less special items); txt / (pi – spi), per Compustat. Set to missing if the denominator is negative.
TaxEvent*WeakRM = Interaction of TaxEvent and WeakRM variables. Control*WeakRM = Interaction of Control and WeakRM variables. Control*StrongRM = Interaction of Control and StrongRM variables. TaxEvent*StrongRM = Interaction of TaxEvent and StrongRM variables. PDTR = Predicted downside tax risk, based on coefficients from the matched-
sample prediction model in Table 5 column 2. For use in holdout sample tests.
PDTR*PrTaxEvent = Interaction of PDTR and PrTaxEvent for use in holdout sample tests. StrongRM*PDTR = Interaction of StrongRM and PDTR for use in holdout sample tests. StrongRM*PrTaxEvent = Interaction of StrongRM and PrTaxEvent for use in holdout sample tests. StrongRM*PDTR* PrTaxEvent
= Interaction of StrongRM, PDTR and PrTaxEvent for use in holdout sample tests.
MKTRF = Excess market return in month t, the value-weighted market return minus risk-free rate [rmt - rft]; mktrf, per CRSP.
SMB = Size premium for month t, the difference in returns bet