Examining which tax rates investors use for equity valuation

Transcription

1 Examining which tax rates investors use for equity valuation Kathleen Powers University of Texas at Austin Jeri Seidman University of Virginia and Bridget Stomberg University of Georgia May 2016 Acknowledgements: We thank John Campbell, Judson Caskey, Novia Chen (discussant), Michael Clement, Lisa De Simone, Ross Jennings, Lisa Koonce, Lillian Mills, Casey Schwab, Brian White, the University of Arizona and the University of Iowa tax readings groups, workshop participants at University of Georgia, the University of Texas at Austin and the University of Waterloo, and participants at the 2014 AAA Annual Meeting for helpful suggestions. Powers gratefully acknowledges financial support from the AICPA Foundation through the Accounting Doctoral Scholars Program and the Red McCombs School of Business.

2 Examining which tax rates investors use for equity valuation ABSTRACT: We propose that investors rely on tax rate heuristics to reduce information processing costs associated with understanding complex income tax information and examine which information investors use to impound income taxes into firm value. We find that tax expense estimated using the top U.S. statutory rate is more associated with firm value than the firm s prior-year effective tax rate (ETR), prior three-year average ETR or prior-year industry-average ETR. However, we find that investors rely less on the statutory tax rate when the benefits (costs) of doing so are higher (lower). Investors incorporate industry-specific tax information more for firms with high future tax planning opportunities. Additionally, investors incorporate firm- and industry-specific information more when information processing costs are expected to be lower. Our findings advance the literature regarding information processing costs, inform the valuation of tax literature and have implications for management in communicating tax information to investors. Keywords: Valuation, Tax expense, Stock returns Data Availability: Data are available from public sources identified in the paper. JEL classification: G12, M40, M41 Page ii

3 I. Introduction We examine which income tax information investors incorporate into firm value. Income taxes are a material expense for U.S. corporations and, as such, should affect valuation. Yet the complexity of the tax code and the rules that govern accounting for income taxes make it difficult for investors to comprehend income tax disclosures and incorporate future tax outcomes into firm value. Indeed, prior literature finds that sophisticated financial statement users struggle to impound anticipated changes in tax expense into estimates of future performance (e.g., Chen and Schoderbek 2000; Plumlee 2003; Weber 2009). Similarly, in concurrent work, Graham, Hanlon, Shevlin and Shroff (2016) provide evidence that managers often use statutory tax rates when making decisions despite having the skills and information necessary to more precisely estimate tax effects. 1 These findings suggest that the cost of processing tax information to develop refined expectations may not outweigh the benefits for investors, analysts and managers when making valuation decisions. We propose that investors reduce their information processing costs by relying on heuristics when impounding taxes into firm value (Payne 1976, 1982). Following Gigerenzer and Gaissmaier (2011), we define a heuristic as a strategy that ignores part of the information, with the goal of making decisions more quickly, frugally and/or accurately than more complex methods. We identify four tax rates heuristics investors can use to estimate tax expense and evaluate which measure is most associated with firm value, on average. The four heuristics are: the firm s prior-year GAAP effective tax rate (ETR), the firm s three-year average GAAP ETR, the average GAAP ETR of the firm s industry, and the top U.S. corporate federal statutory rate of 1 On average, across seven types of corporate decisions including M&A and capital structure decisions, public (private) firm managers report using the following rates to incorporate taxes into forecasts or decision making processes: U.S. Statutory rate: 20% (34.1%); GAAP ETR: 27.4% (20.5%); Jurisdiction-specific statutory tax rate: 21.0% (14.6%); Jurisdiction specific ETR: 17.6% (15.0%); Marginal tax rate: 10.8% (12.5%). Page 1

4 35 percent. Each rate has strengths and weaknesses in estimating tax expense, and so we make no prediction regarding which measures of tax expense are associated with firm value, on average. 2 Our research methodology is similar in spirit to Francis, Schipper and Vincent (2003) who examine the relative and incremental explanatory power of various earnings measures to provide evidence about aggregate investor behavior for valuation. To determine if a measure of tax expense calculated using a tax rate heuristic is associated with firm value, we estimate annual cross-sectional regressions of 12-month contemporaneous buy-and-hold returns from 1996 through 2013 as a function of pre-tax earnings surprise, tax surprise, and controls. The valuation of tax literature commonly assumes tax expense follows a random walk similar to earnings (e.g., Ayers, Jiang and Laplante, 2009; Hanlon, Laplante and Shevlin, 2005; Thomas and Zhang, 2014). These models therefore implicitly assume both the GAAP ETR and pre-tax income follow a random walk (because tax expense is ETR multiplied by pretax income). Consistent with this literature, we allow ETRs (and therefore tax expense) to follow a random walk by including the difference between current and prior-year tax expense as a measure of tax surprise ( PY tax surprise ). We also modify this model to include other tax surprise, which represents the difference between prior year tax expense and expected current period tax expense estimated using the remaining tax rate heuristics that we examine (i.e., firm s three-year average GAAP ETR, the firm s industry average GAAP ETR and the U.S. top statutory rate of 35 percent, respectively). This design allows us to determine whether other tax information is incrementally informative to investors when impounding taxes into firm value. 2 Additionally, we acknowledge that the tax rates we examine are not an exhaustive set of tax information available to investors when making valuation decisions. It is therefore possible that our results reflect investors use of information that is highly correlated with, but different from, our chosen measures. For example, investors might use a 5-year firm-average ETR, which we expect would be highly correlated with our three-year measure. Page 2

5 A significant coefficient on PY tax surprise suggests investors use a random walk to value current year tax expense. A significant coefficient on other tax surprise suggests tax information other than that reported in the prior period is incrementally informative to investors when impounding taxes into value. Following the methodology in Francis et al. (2003), we test the relative information content of each model by: (1) counting the number of annual regressions for which we estimate a statistically significant coefficient on PY or other tax surprise and (2) using a Vuong (1989) test to count the number of years in which a particular model has the highest (is tied with another model for the highest) adjusted R 2. On average, we estimate statistically significant coefficients on other tax surprise in nearly every year for models using either the top U.S. statutory tax rate or the industry average rate to estimate tax expense. The statutory model generates the largest adjusted R 2 in as many as 18 of the 18 years, depending on the specification, whereas the industry model generates the largest adjusted R 2 or is tied with the statutory model in at most four years. We therefore conclude that investors most often impound taxes into firm value using the statutory tax rate, on average in a broad sample of firms. This result compliments recent survey evidence by Graham et al. (2016), as well as experimental evidence by Amberger, Eberhartinger, and Kasper (2016), that individuals often use statutory tax rates instead of firm-specific tax information when making decisions. Although the coefficient on firm average surprise is significant in as many as 16 of the 18 years, the firm average model does not generate the largest adjusted R 2 in any year. PY tax surprise also does not generate the largest adjusted R 2 in any year, even when we restrict the sample to observations with historically low ETR volatility, where we expect prior-year information to be most useful. Finding that the statutory model dominates the PY tax surprise model is potentially Page 3

6 surprising given that prior literature often specifies tax surprise relative to the prior-year and almost never relative to the statutory rate. We further explore these results below. The statutory rate is an extremely low cost heuristic. However, Dyreng, Hanlon, Maydew and Thornock (2015) find that reported ETRs can be significantly lower than 35 percent, such that if investors use the statutory tax rate to impound taxes, they may overestimate tax costs and underestimate firm value. We conduct analyses to investigate whether investors reliance on the statutory rate is reasonable and whether it changes when benefits (costs) of using a different heuristic are higher (lower). First, we document that for the median profitable firm, ETR converges to 35 percent after only four years, suggesting that the costs of using the statutory tax rate may not outweigh the benefits of developing a more refined model. We further test whether investors rely less on the statutory tax rate when the benefits of incorporating additional information are presumably higher. We re-estimate our regressions on subsamples of firms with potential opportunities for future tax avoidance evidenced by either high levels of research and development (R&D) expenditures and foreign sales, or prior tax avoidance (i.e., firms with historically low industry-adjusted ETRs). Greater potential opportunities for future tax savings should result in larger errors if investors use the statutory rate to impound taxes. We continue to find that the statutory model generates the highest number of statistically significant coefficients and the highest explanatory power in the most number of years in these two subsamples. However, in these subsamples we estimate an indistinguishable difference between the explanatory power of the statutory and industry models in eight (seven) of 18 years. This suggests that though investors continue to focus on the statutory tax rate when valuing taxes for firms with greater opportunities for long-term tax planning (evidence of successful tax planning), they incorporate industry-specific information more often than in the full sample of firms. We also Page 4

7 find some evidence of investors relying more on firm-specific information in this subsample relative to the full sample. Second, we examine whether investors use different tax rates when information processing costs are lower. Though each of the heuristics is relatively easy to calculate, investors still must devote time to understanding whether the heuristic yields a reasonable estimate of tax expense. We measure information processing costs in two ways. First, we use the presence of analyst coverage as an indication of a richer information environment and hence, lower investor processing costs. Analysts can reduce investor processing costs by gathering, summarizing, and interpreting a broader set of both firm- and industry- specific information for investors (Healy and Palepu 2001). Second, we consider more sophisticated investors (i.e., institutional owners) to be better able to process complex tax information (Blankenspoor 2015; Dye 1998; Fishman and Hagerty 2003). We find evidence that investors incorporate firm- and industry-specific tax information more when firms have active analyst coverage and when firms are in the top quintile of institutional ownership. Specifically, we find that the industry model estimates a statistically significant coefficient in more years than does the statutory model and that the industry model generates the statistically highest R 2 at least as frequently as does the statutory model. Additionally, we find much more reliance on firm-specific tax information relative to the full sample of firms. Taken together with the previous set of cross-sectional tests, we conclude that investors incorporate firm- and industry-specific tax information into their valuation decisions more when the benefits (costs) of doing so are higher (lower). Our study complements and extends the literature on how capital market participants use tax information (Ayers et al. 2009; Chen and Schoderbek 2000; Hanlon et al. 2005; Plumlee 2003; Schmidt 2006; Thomas and Zhang 2014) by examining which tax information investors impound Page 5

8 into firm value. Whereas much prior literature assumes that investors use a firm s prior-year tax expense to set expectations about future taxes, our results suggest the U.S. statutory tax rate is most associated with firm value. We also contribute to the literature that analyzes how various stakeholders incorporate taxes into investment decisions. Our finding that the statutory tax rate is most associated with firm value reveals that investors, as stakeholders outside the firm, use the simplest heuristic to impound taxes. These results are consistent with survey evidence from Graham et al. (2016) that corporate managers, as stakeholders inside the firm, often rely on the statutory tax rate when making investment decisions. Finally, we contribute to the literature documenting the effects of information processing costs on investors use of financial statement information. Consistent with prior literature (Hong, Lim, and Stein 2000; Soffer and Lys 1999; Walther 1997), we find that in a richer information environment, investors incorporate more industry-specific information into firm value. Our results also have several implications for managers who seek to understand how investors use financial statement information and incorporate this information into price. Our finding that investors rely heavily on the statutory tax rate, and that they tend to supplement with the industry average rate when they seek additional tax information, implies they often ignore firmspecific tax information when making valuation decisions. This finding is consistent with prior research documenting investors limited attention for understanding and absorbing financial statement information (Daniel, Hirshleifer, and Teoh 2002; Hirshleifer, Lim, and Teoh 2009). To decrease investors information processing costs, managers can focus discussions on persistent differences between the firm s ETR and the statutory tax rate or industry average rate, which may allow better assimilation of more relevant firm-specific information into price. Page 6

9 II. Background and Prior Literature Overview Income taxes are a material and recurring expense for most U.S. corporations and are therefore an important component of firm value. From 1996 through 2013, the average (median) profitable Compustat firm reported income tax expense equal to 25.4% (32.3%) of pre-tax income, 3.6% (2.7%) of sales, and 2.8% (2.4%) of market capitalization. In contrast, R&D, which is often considered important for valuation purposes, is only 2.1% (0.0%) of sales and 1.4% (0.0%) of market capitalization. Tax expense is also a larger percentage of sales than either interest expense or advertising expense. 3 These statistics reveal not only that income taxes are of sufficient magnitude to warrant investors consideration in valuation decisions but that they are perhaps one of the most significant expenses to consider. Prior studies provide evidence that income taxes are value relevant. Lev and Thiagarajan (1993) identify income taxes as one firm fundamental that explains equity prices. However, recent literature on tax-law and tax-disclosure changes demonstrates that income taxes are difficult for even sophisticated users to comprehend. Chen and Schoderbek (2000) find that analysts failed to properly adjust their earnings forecasts to include the effect of the statutory tax rate change from the 1993 Omnibus Budget Reconciliation Act on the deferred tax accounts even though the information required to estimate the one-time expense or benefit was available. Similarly, Plumlee (2003) examines the tax-law changes resulting from the Tax Reform Act of 1986 and finds that analysts incorporated the effect of less complicated tax law changes into their earnings forecast 3 Statistics in this paragraph are based on 91,585 Compustat observations from 1996 to 2013 where pre-tax income (PI), sales (SALE) and market capitalization (PRCC_F*CSHO) are all greater than 0. Research and development (XRD), advertising (XAD) and interest expense (XINT) are all set to 0 if missing. All variables are scaled by either sales or market capitalization and ratios are winsorized by year at one and 99 percent to avoid inflated averages resulting from small denominators. Page 7

10 but did not incorporate the effect of more complicated changes. Even when no tax law changes occur, Bratten, Gleason, Larocque and Mills (2015) and Kim, Schmidt and Wentland (2015) find that analysts struggle to properly incorporate tax information into earnings forecasts. Regarding tax disclosure changes, Robinson, Stomberg and Towery (2015) find no evidence that investors can identify firms with tax reserves that are most likely to be settled in cash based on tax reserve disclosures. Indeed, it remains unclear whether or when changes in tax expense are positively or negatively associated with returns. Hanlon et al. (2005) document that taxable income estimated from financial statement tax expense is positively related to returns and is incrementally informative to pre-tax income in explaining returns. The authors posit that taxable income is an alternative measure of firm performance. Building on the idea that income tax expense can serve as a proxy for economic profitability, Ayers et al. (2009) find that investors rely more (less) on taxable income as an alternative performance measure when earnings quality is low (tax planning is high). However, Thomas and Zhang (2014) show that tax expense is informative about future profitability only in model specifications that do not otherwise control for estimated future performance. In samples where earnings surprises are small and in specifications that include controls for expected future profitability, they find that income tax expense is valued as a cost that represents value lost to tax authorities (i.e., is negatively associated with returns). These results collectively suggest that the cost of processing firms income tax information to develop a tax forecasting model is not trivial for the average investor. This assertion is further supported by the fact that the majority of firms tax information is contained in footnote disclosures, which have been shown in experimental studies to increase processing cost (e.g., Hirst and Hopkins 1998). Thus, we propose that investors rely on heuristics to incorporate tax Page 8

11 information into firm value instead of trying to understand the drivers and characteristics of firms tax expense. The psychology literature demonstrates that people rely on simple decision mechanisms to help them process complex information. Payne (1976) finds that as task complexity increases, participants resort to limited-information decision strategies that allow them to eliminate some options as quickly as possible. One such simple decision mechanism is a heuristic, which is a strategy that ignores part of the information, with the goal of making decisions more quickly, frugally, and/ or accurately than more complex methods (Gigerenzer and Gaissmaier 2011). In concurrent work, Graham et al. (2016) present evidence consistent with corporate managers using statutory tax rates as a heuristic when evaluating investing, financing and operating decisions despite the availability of relevant firm-specific information that would lead to more accurate estimations. We examine which of four tax rate heuristics investors use to impound taxes into firm value and, in doing so, identify which tax information investors may be ignoring in their valuation decisions. Tax rates as heuristics We use four tax rates in our analysis, each of which meets the definition of a heuristic because it allows the user to ignore some relevant information with the goal of making decisions more quickly than more complex methods. The four rates are: (1) the highest corporate U.S. statutory tax rate (Stat_Rate), (2) the firm s prior-year ETR (PY_ETR), (3) the average of the firm s three prior annual ETRs (FirmAvg_ETR), and (4) the prior-year industry-average ETR (IndAvg_ETR). 4 All four rates are relatively easy for investors to obtain. We discuss our motivation 4 We limit the rates we examine to those that are available to the general public and acknowledge that our list is not exhaustive. Page 9

12 for selecting these rates below, beginning with the heuristic with the lowest processing cost and ending with heuristics that are more difficult to calculate or to understand. Stat_Rate, equal to 35 percent since 1993, is readily-available to investors because all public corporations must reconcile their ETR to the U.S. corporate statutory tax rate in their income tax footnote. The business press also frequently uses the U.S. statutory rate as a benchmark against which to evaluate firms taxes. For example, in conjunction with Apple s testimony before the U.S. Senate, Bloomberg noted that the company s 30.5 percent tax rate in the U.S. lags behind the corporate tax rate of 35 percent (Drucker 2013). Articles such as these make the statutory tax rate very salient to investors and perhaps give the impression that this is the rate investors should be using when evaluating corporate taxes. Evidence also suggests that analysts view reductions in tax expense not caused by changes in the statutory tax rate as transitory (Abarbanell and Bushee 1997, 1998; Lev and Thiagarajan 1993), suggesting these deviations are down-weighted in valuation decisions. Finally, the statutory tax rate is potentially a useful heuristic because many tax reduction strategies are temporary in nature. Thus, benefits claimed in one period reverse in a future period such that corporations will pay an average rate of 35 percent to the U.S. on all pretax earnings over their lifetime, absent permanent tax reduction strategies and ignoring the timevalue-of-money. However, using the statutory rate to estimate future tax outcomes ignores the fact that many tax planning strategies allow significant deferral of tax payments. Schmidt (2006) notes that many reductions in tax expense relative to the statutory tax rate reflect long-term (and therefore persistent) strategic tax planning. These strategies include transfer pricing, tax-efficient supply chain management and tax-favored intercompany debt structures. Therefore, although using the statutory tax rate to impound taxes for valuation is low cost and may be valid over very long Page 10

13 periods, it ignores information about firm- or industry-specific opportunities to generate valuerelevant tax savings and may result in investors overestimating tax costs in the short term. Using a firm-specific tax rate overcomes this limitation. Corporations must present information about both their current and prior-year taxes in the income tax footnote of the annual report and frequently discuss differences between the two in the MD&A. Prior-year tax expense and pre-tax income are also presented on the face of the income statement, allowing investors to easily calculate this ratio without relying on footnote disclosures. Thus, PY_ETR (calculated as TXTt-1/PIt-1) is salient to investors and of relatively low cost to obtain. It also has the advantage of containing the most recent information about firm-specific characteristics that contribute to a firm s ability and willingness to avoid tax. Although PY_ETR has certain information advantages, it can also be noisy due to periodic settlements with taxing authorities, significant one-time corporate transactions, and earnings management through the tax accrual (Dhaliwal, Gleason and Mills 2004; Robinson et al. 2015). Understanding whether a firm s prior-year ETR is likely to generate a reasonable estimate of future taxes is not costless because it requires investors to understand which components of tax expense are persistent and which are transitory. Raedy, Seidman and Shackelford (2012) document the difficulty inherent in understanding the income tax footnote through a collection of detailed footnote data. The examples of rate reconciling items they provide demonstrate the inconsistent language firms use to refer to similar underlying transactions. They further find that approximately 90 percent of the rate reconciliations they collect include a line entitled Other, Miscellaneous, or a similarly vague description that gives no information on the underlying transactions (Raedy et al. 2012). The authors state, During the process of collecting, interpreting and categorizing the information, we were repeatedly struck by the difficulty in understanding these data. Therefore, Page 11

14 we expect even sophisticated investors experience difficulty in processing the information underlying firms ETRs. To gain a better understanding of which deviations from the statutory tax rate are persistent rather than transitory and to reduce the noise in the one-year measure, market participants can average tax expense over multiple periods to arrive at a firm-average ETR (e.g., Dyreng, Hanlon, and Maydew 2008). This heuristic will reduce the effect of significant one-time deviations from normal trends. However, because firms often provide information for only three years, calculating a longer-horizon ETR will require effort on the part of investors. Long-run averages can also mask informative tax volatility. Keeping with the spirit of a heuristic, we use only the information included in a single annual report and calculate FirmAvg_ETR over a three-year window. Thus, FirmAvg_ETR is the average ETR from t-3 through t-1 where ETR is defined as (TXT/PI). Industry-wide tax rates can provide relevant information about a firm s taxes because firms within the same industry tend to have comparable income tax avoidance opportunities and are oftentimes similarly affected by changes in tax legislation (e.g., Balakrishnan, Blouin, and Guay 2012; De Simone, Stomberg, and Mills 2015). Consistent with this notion, prior studies show that analysts and investors frequently use industry-average performance to set expectations about and evaluate firm-specific performance (Lev 1989). Industry-average tax rates can therefore provide information about potential changes or trends that are not yet reflected in a particular firm s tax expense. However, this rate is more difficult for investors to calculate because it requires cumulating information across multiple companies and knowing at what level to define an industry. 5 We include IndAvg_ETR to capture the average ratio of tax expense to pre-tax income 5 To be useful as a tax rate heuristic, an industry-average ETR must effectively combine firms with similar opportunities for income tax avoidance. Common industry definitions include one-, two-, or three-digit SIC codes, as well as other groupings based on 4-digit codes such as those provided by Fama and French. GCIS and NAICS codes Page 12

15 in a given industry and calculate IndAvg_ETR as the firm s industry average ETR in year t-1 where industry is defined using the Fama-French 30 industry classifications. To summarize, we propose that investors weigh the relative advantages and disadvantages of each of the tax rates discussed above when deciding how to impound taxes into firm value. Therefore, it is not clear, ex ante, which calculation of tax expense is most associated with firm value. Thus, we make no predictions about which rate(s) investors use and instead examine this question empirically. III. Research Method Regression Specification Much prior research examining the association between taxes and firm value assumes that both pre-tax income and tax expense follow a random walk. These studies therefore measure surprises using one-year changes in pre-tax income and tax expense (e.g., Lev and Thiagarajan 1993; Thomas and Zhang 2014). 6 To determine whether other tax information is relevant to investors when impounding taxes into price, we estimate the following specification using annual cross-sectional regressions: Retit =β0 +β1py_tax_surpriseit + β2other_tax_surpriseit +β3income_surpriseit +β4logmveit-1 +β5retit-1 +β6btmit-1 +εt (1) where Retit is the buy-and-hold return to security i over a 12-month window beginning at the end of the third month of year t and ending at the end of the third month of year t+1. Our approach is motivated by Francis et al. (2003), who examine the association between long-window returns and are also commonly used. For example, Lev (1989) classifies industry using two-digit SIC codes while Balakrishnan et al. (2012) and Dyreng et al. (2008) use the Fama and French 30 (FF30) classification. 6 Some studies use analysts consensus earnings forecasts as a proxy for investors expectations. However, forecasts of pre-tax earnings are not well populated in IBES. According to Mauler (2015), only 1,031 (3,500) pre-tax earnings forecasts are available in IBES in 2002 (2013). Thus, relying on IBES for pre-tax income forecasts would reduce our sample by up to 65 percent in some years. Our analysis focuses on the relative explanatory power of the models we test and holds the value of PI_Surprise constant across models. Thus, we have no reason to believe that our results depend on the measure of PI_Surprise we use. However, we explore analysts ETR estimates in section VI. Page 13

16 various measures of firm performance to provide evidence about aggregate investor behavior. We follow prior studies that allow tax expense to follow a random walk and calculate PY_Tax_Surpriseit as follows: PY_Tax_Surpriseit = (TXTt TXTt-1) / MVEt (2) where MVE is the market value of equity three months after the end of year t. This definition of PY_Tax_Surpriseit implicitly assumes ETRs follow a random walk and allows us to test the prioryear effective tax rate (PY_ETR) as a heuristic investors use to value current period tax expense. We calculate PY_Tax_Surpriseit such that if actual tax expense is lower than prior-year tax expense, the value is negative. Prior studies document that investors view tax expense either as representing value lost to tax authorities or as a proxy for future profitability (Ayers et al. 2009; Thomas and Zhang 2014). Although an investigation of the differing roles of tax expense is not the focus of this study, in our tests, a negative coefficient on PY_Tax_Surprise (β1 < 0) is consistent with taxes being viewed as an expense, or value lost (e.g., Lipe 1986). Conversely, β1 > 0 is consistent with taxes serving as a proxy for profitability in our sample. Because of these conflicting suppositions, we make no prediction as to the sign of β1. 7 We evaluate whether investors incorporate information other than prior-year tax expense by including a second measure of tax surprise. Other_Tax_Surpriseit captures the difference between prior-year tax expense and expected current year tax expense calculated by multiplying each heuristic (other than PY_ETR) in turn, by PIt-1: Other_Tax_Surpriseit = (TXTt-1 TROTHER * PIt-1) / MVEt (3) 7 A potential concern is that because of these competing roles for tax expense, leading to opposite signs, a heuristic may be used for both roles leading to an insignificant coefficient, on average. To mitigate this concern, we follow Thomas and Zhang (2014) and present results both before and after trimming extreme values of PI_Surprise. Page 14

17 where TROTHER is Stat_Rate, FirmAvg_ETR or IndAvg_ETR. Other_Tax_Surpriseit represents the difference between tax expense predicted using a random walk and tax expense predicted using one of our three alternative tax rate heuristics. Thus, if prior-year tax expense is lower than tax expense predicted using the heuristic, the value is negative. When we evaluate PY_ETR as a heuristic, Other_Tax_Surprise equals zero. Estimating β2 < 0 is consistent with an increase in expected taxes representing value lost while estimating β2 > 0 is consistent with increased tax expense signifying future profitability. For all heuristics, the sum of PY_Tax_Surpriseit and Other_Tax_Surpriseit captures the total difference between reported tax expense and expected tax expense calculated using each of the benchmarks and PIt-1. Estimating a significant coefficient on Tax_Surprise suggests that particular tax rate heuristic contains information relevant for investors valuation. Following Francis et al. (2003), we determine if one particular model includes a tax expense that is more closely aligned with the information investors use to impound taxes into firm value than the other models in two ways. First, we count the number of years in which the coefficient on each measure of Tax_Surprise is statistically significant. A model with a higher number of years of coefficients different than zero is considered more associated with firm value. Second, we use a Vuong (1989) z-statistic to test whether the explanatory power of any model is significantly higher than the explanatory power of all other models each year. We consider the model that dominates all others in the greatest number of years to contain the calculation of tax expense investors most closely associate with firm value. Insignificant differences in explanatory power between the models are consistent with no investor preference for a particular tax rate, on average. Page 15

18 We include several controls in our model. To ensure our measure of tax surprise is not simply picking up changes in profitability, we control for pre-tax income surprise (PI_Surprise) as follows: PI_Surpriseit = (PIt PIt-1) / MVEt (4) and expect β3 >0. Additionally, following Thomas and Zhang (2014), we include controls for other determinants of observed returns including the natural log of the market value of equity and bookto-market ratio at the end of year t-1, as well as returns for the prior-year s 12-month period with a one-month lag relative to Rett. Prior literature suggests returns to be decreasing in logged, lagged market value (β4 <0) and in prior-year returns (β5< 0), and increasing in lagged book-to-market ratio (β6 >0). Henceforth, we omit firm and year subscripts for simplicity. Sample We begin our sample by selecting all firm-years in the intersection of Compustat and CRSP from 1993 through 2013 with data necessary to calculate the required variables. Consistent with Thomas and Zhang (2014), we do not eliminate observations reporting pre-tax losses or observations with extreme ETRs (i.e., less than zero or greater than one) but do winsorize all variables except Ret at the top and bottom one percent. We first estimate equation (1) on our full sample of winsorized observations. To assess how observations with extreme income surprises affect our results, we also estimate equation (1) on two subsamples where we trim observations based on PI_Surprise. Thomas and Zhang (2014) argue that this step-wise approach strengthens the relation between pre-tax income and future profitability, thereby allowing tax expense to better represent value lost to tax authorities and serve less as a proxy-for-profitability. This approach also allows us to test whether investors use different Page 16

19 heuristics in situations where tax surprise might be heavily influenced by extreme changes in profitability. Table 1 details our sample selection process. We require observations to have sufficient data to calculate each of the tax rate heuristics we examine so that differing results across model specifications are not due to changes in sample composition. Because we want all observations in our sample to account for income taxes consistent with ASC 740, we begin the calculation of FirmAvg_ETR, which requires three years of data, in Thus, our regression period is 1996 through To calculate a meaningful industry-average ETR, we require each industry to have at least 10 observations per year. Our final sample consists of 86,310 firm-year observations from 11,503 unique firms. Firms are in our sample for 7.5 years, on average. [Insert Table 1 here.] IV. Results Descriptive Statistics Panel A of Table 2 reports descriptive statistics for GAAP ETRs as well as for the tax rate heuristics we test. We compute GAAP ETR (ETR) as total tax expense scaled by pre-tax income (TXT/(PI)). The average (median) ETR for firms in our sample is 19.7 (29.2) percent. The U.S. statutory rate is 35 percent for every year in the sample. The average values for the other ETR heuristics are 20.4 percent for PY_ETR, 20.6 percent for FirmAvg_ETR, and 24.4 percent for IndAvg_ETR. [Insert Table 2 here.] We scale all of our tax and pre-tax earnings surprise variables by market value of equity when we estimate regressions. We present descriptive statistics on scaled regression variables in Page 17

20 Panel C. 8 For ease of interpretation, we report descriptive statistics for unscaled values (in $M) in Panel B. The average unscaled value of PI_Surprise is $12.63M, indicating that firms report a year-over-year increase in pre-tax earnings, on average. PI_Surprise is also positive at the median. PY_Tax_Surprise is $3.36 on average, indicating that increases in pre-tax income generate increases in total tax expense. Stat_Surprise is negative, on average, consistent with descriptive statistics in Panel A showing that, on average, reported ETRs are less than 35 percent. FirmAvg_Surprise and IndAvg_Surprise are both positive at the mean and median. Descriptive statistics on the scalar, market value of equity, show that the average observation in our sample is large, with market capitalization of over $2B. In Panel C, we report that the average (median) annual buy-and-hold return is 18.7 (6.2) percent in our sample. Table 3 provides Pearson and Spearman correlations among regression variables. Pearson correlations are listed above the diagonal with Spearman correlations below. Most Tax_Surprise variables are significantly correlated with two-tailed p-values 10 percent. However, untabulated VIF scores indicate that multicollinearity is not a significant concern. [Insert Table 3 here.] Main Analysis Table 4 presents results of estimating equation (1) as annual cross-sectional regressions. We present results in a stepwise manner by first estimating returns as a function of total net income surprise. This model allows total net earnings to assume a random walk model for expected future income. We then disaggregated the total change in net income and estimate equation (1) annually substituting each of the four tax rate benchmarks, in turn, and present the average coefficients and average adjusted R 2. 8 Untabulated T-tests indicate that all measures of Tax_Surprise are statistically different. Page 18

21 Panel A presents results of estimating equation (1) on our full sample. Following Thomas and Zhang (2014), we also trim the sample based on extreme values of PI_Surprise to increase the likelihood that tax expense represents value lost to tax authorities and is not a proxy for profitability. In Panel B, we report results after trimming the top and bottom five percent of PI_Surprise, and Panel C presents results after trimming the top and bottom ten percent of PI_Surprise. Because the models offer the highest explanatory power in Panel B, we focus our discussion of results on this sample. We also use this as our baseline specification in all crosssectional tests in Section V. [Insert Table 4 here.] In column (1) we estimate returns as a function of NI_Surprise. We estimate a positive coefficient on NI_Surprise, as expected. The average R 2 for this model appears relatively low and does not increase appreciably in column (2) when we decompose NI_Surpise into its pre-tax income and tax expense components. Thus, changes in current-year tax expense relative to prioryear tax expense do not appear to contribute a large amount of explanatory power for returns. In all columns of Panel B, the coefficient on PI_Surprise is positive and significant, and the sign and magnitude of coefficients on control variables are as predicted. We estimate negative and significant average coefficients on PY_Tax_Surprise in all four columns, which implies that firms reporting decreases (increases) in tax expense have greater (lower) returns, all else equal. 9 We test whether other tax information has explanatory power for returns in columns (3) through (5), where we include Other_Tax_Surprise. In all three columns, we estimate negative and significant average coefficients on Other_Tax_Surprise, indicating that information contained 9 Comparing across the three panels, though we estimate an insignificant coefficient on PY_Tax_Surprise in two of the four columns in Panel A, we estimate significantly negative coefficients on PY_Tax_Suprise in all four columns of both Panels B and C. This is consistent with tax expense taking a weaker role as a proxy for profitability when PI_Surprise is less extreme. Page 19

22 in the measures of tax expense calculated using other tax rate heuristics is incrementally informative to one-year changes in tax expense when explaining cross-sectional variation in returns. 10 We tabulate the number of years that the coefficient of interest PY_Tax_Surprise in Column (2) and Other_Tax_Surprise in Columns (3) through (5) is significant. The statutory model and industry model produce a significant coefficient on Other_Tax_Surprise in 17 of the 18 years while the PY model and firm average model generate significant coefficients in only five and ten of the 18 years, respectively. The average explanatory power of the statutory and industry models also appears much larger than the other models. We test the relative explanatory power of each model annually using Vuong (1989) tests and find that the statutory model produces the largest R 2 in 14 of the 18 years in our sample. 11 The industry model produces the largest R 2 in two years and in the remaining two years, the explanatory power of the statutory model and the industry model are statistically equivalent and significantly higher than either firm-specific model. Thus, results in Table 4 suggest that tax expense estimated using the U.S. statutory rate is more associated with firm value than that calculated using IndAvg_ETR. Additionally, tax expense calculated using IndAvg_ETR is more associated with firm value than that calculated using either PY_ETR or FirmAvg_ETR. V. Cross-sectional Tests Our results suggest tax expense calculated using either the statutory tax rate or an industryspecific tax rate is more associated with firm value than tax expense calculated using firm-specific tax rates. This result is surprising because prior literature (e.g., Ball and Watts 1972; Beaver 1970; 10 Further consistent with tax expense taking a weaker role as a proxy for profitability when PI_Surprise is less extreme, the coefficient on the various specifications of Other_Tax_Surprise generally becomes more negative across the panels as we trim observations with extreme pre-tax earnings surprises. 11 Comparing across the panels, the statutory model continues to yield the highest average adjusted R 2 as we trim observations with more extreme PI_Surprise. However, the number of years in which the statutory model generates the highest annual adjusted R 2 falls from all 18 years in Panel A to 14 and 11 years in Panels B and C, respectively. Page 20

23 Watts and Leftwich 1988) documents that net income follows a random walk or a random walk with a drift. As such, tax researchers (e.g., Ayers et al. 2009; Hanlon et al. 2005; Thomas and Zhang 2014) generally calculate tax surprise as the year-over-year change in tax expense, which implicitly assumes that tax expense also follows a random walk. We address these potentially surprising results in two ways. First, we provide descriptive statistics on long-run ETRs to see if using the statutory tax rate to impound taxes into firm value appears reasonable, on average. Second, we examine if investors incorporate firm- or industry-specific information more when the expected benefits (costs) of doing so are higher (lower). Although using the statutory tax rate is possibly the lowest cost way to impound taxes into firm value, the cost-benefit trade-off of using this heuristic likely varies with firm and investor characteristics. Therefore, we expect the benefits of incorporating firm- or industry-specific information are higher when firms have greater potential opportunities for long-run tax avoidance. We expect the costs of incorporating firm- or industry-specific information are lower when information processing costs are lower. Time-trends in Long-Window ETRs We first examine time-series trends in long-window ETRs for a broad sample of firms. Figure 1 plots rolling ETRs over periods of one to 20 years. We calculate these ETRs as the sum of total tax expense (TXT) scaled by the sum of pre-tax income (PI) over various time periods. We restrict this analysis to observations where the sum of pre-tax income over each specified time interval is positive to ease interpretation and mitigate the effects of transitory losses. We observe that the median ETR averages 35 percent after four years and continues to average 35 percent for the remainder of the windows we estimate. Untabulated results are consistent if we restrict the sample to firms with sufficient data available to calculate a 20-year ETR, without regard to overall profitability. It therefore appears that for the median profitable firm, 35 percent is a reasonable Page 21

24 estimation of long-run tax effects. This gives some comfort to results that suggest tax expense estimated using the statutory rate is more associated with firm value than tax expense estimated using an industry- or firm-specific effective tax rate. [Insert Figure 1 here.] Tax Planning Dyreng et al. (2008) report that some firms sustain low cash ETRs for periods of up to ten years. Thus, some firms are able to enhance firm value through strategic tax planning to a greater extent than others. We therefore examine whether investors impound taxes into firm value using different tax rate heuristics when firms have either demonstrated significant historical tax planning or when firms have characteristics associated with opportunities for tax savings. Both a history of tax avoidance and the availability of tax avoidance opportunities can indicate potential future tax savings. Our first proxy for potential future tax savings is an ex post measure based on ETR realizations. Balakrishnan, Blouin and Guay (2012) propose using an industry- and size-adjusted ETR to capture aggressive or unexpected tax avoidance. They posit that all else equal, firms in similar industries and of similar size have similar tax planning opportunities. Thus, ETR realizations lower than the industry-size average ETR indicate firms have undertaken additional tax minimizing strategies in that year. To identify firms who have achieved more significant tax savings, we follow Balakrishanan et al. (2012) and compute a three-year industry-size adjusted GAAP ETR, which is the difference between the industry-size average ETR and firm s ETR from t-3 to t-. The measure is constructed so that positive values represent tax avoidance in excess of industry-size peers. We consider observations in the top quintile of this adjusted ETR to have greater long-term tax avoidance. A strength of this measure is that it does not require us to identify Page 22

25 specific tax minimizing strategies. However, a disadvantage is that we could misclassify firms because GAAP ETRs reflect both real tax planning and financial accounting decisions. For example, tax contingency reserves can mask the extent of tax avoidance. To mitigate this disadvantage, we also identify firms with high levels of R&D and foreign sales because these characteristics are associated with opportunities for long-term tax avoidance. Claiming R&D tax credits permanently reduces taxes and therefore lowers a firm s reported ETR. Indeed, Dyreng et al. (2008) report that firms with the most significant long-run tax savings report larger amounts of R&D expenditures than firms reporting only moderate savings. Similarly, an extensive presence in low-tax foreign jurisdictions allows companies to minimize current taxes by deferring U.S. tax on qualified foreign earnings. Firms can additionally reduce the incremental U.S. taxes due upon repatriation of these earnings through strategic planning, such as tax-efficient supply chains that locate high return activities in low-tax jurisdictions. Using the statutory tax rate exclusively to impound taxes for these firms may therefore result in under-valuation. To identify firms with greater opportunities for long-term tax savings, we independently rank firm-years into quintiles of R&D expense and percent of foreign sales by year. For each ranking, we assign an observation a score of one if it is in the lowest quintile and five if it is in the highest quintile. We then sum the two ranks such that each observation can earn a score ranging from two to 10. Observations in the top quintile of this composite score are deemed to have greater tax planning opportunities. We present the average coefficients and average adjusted R 2 from estimating equation (1) using these two subsamples in Table 5. For comparison purposes, Panel A repeats results from Panel B of Table 4 where we trim the full sample at the top and bottom five percent of PI_Surprise. Panel B of Table 5 presents results where the subsample is defined using industry-size adjusted Page 23