Long-term Payoffs to Aggressiveness

Transcription

1 Long-term Payoffs to Aggressiveness Frank Ecker, Jennifer Francis*, Per Olsson and Katherine Schipper Duke University We examine several long-term consequences to shareholders and CEOs of firms characterized by aggressive accounting during 1974 to Our proxy for aggressiveness is Bowen et al. s (1995) accounting choice score. We provide evidence for the construct validity of this proxy by showing that the most aggressive firms are more likely to restate their financial reports, to be subject to a SEC accounting and auditing enforcement release (AAER), and to report larger unrecognized tax benefits pursuant to FIN 48. We provide evidence that the most aggressive firms have larger cumulative shareholder returns and pay their CEOs more than the least aggressive firms over several long-term horizons. These differences in returns to investors and payouts to CEOs persist after controlling for known determinants of stock returns and compensation, respectively. When we further condition on the existence of a GAAP violation (such as a restatement or an AAER), we find that the long-term stock returns of negative outcome firms are larger than the returns of the remaining most aggressive firms. This difference in returns persists until more than two years after the negative outcome. Our findings suggest that over long horizons, both investors and CEOs of firms characterized by more aggressive accounting benefit, taking into account the effects of adverse reporting outcomes such as restatements and SEC AAERs. September 2011 Preliminary and incomplete; please do not distribute without authors permission * Contact author: Jennifer Francis, jfrancis@duke.edu, Fuqua School of Business, Duke University, Durham, NC,

2 Long-Term Payoffs to Aggressiveness 1. Introduction This study investigates the returns that accrue over long horizons to shareholders and CEOs of firms characterized by aggressive accounting. Our starting point is prior research that reports sharply negative short-window stock market reactions to adverse outcomes associated with aggressive accounting practices, such as announcements of restatements or SEC enforcement actions (for example, Palmrose, Richardson and Scholz [2004], Hennes, Leone and Miller [2008] and Dechow, Sloan and Sweeney [1996]). In showing that investors suffer immediate losses associated with adverse accounting outcomes, this research does not take account of long-horizon returns to those investors in periods leading up to the adverse outcomes. In addition, an adverse outcome has relatively low ex-ante probability of occurring. Our aim is to shed light on the nature of longer-term (that is, multi-year) payoffs associated with aggressive accounting, both unconditionally and adjusting for negative price responses to specified adverse outcomes. Our analysis considers a broad sample of ex-ante aggressive-accounting firms, not just firms that experience adverse outcomes; we evaluate returns over long horizons (3, 5, 10 and 15 years); and we consider payoffs to both equity investors and CEOs. Our research design requires a simple proxy for aggressive accounting that can be calculated for broad samples of firms over long time periods. Relying on previous research, we use Bowen, DuCharme and Shores [1995] composite score of income-increasing or income-decreasing inventory cost-flow assumptions and depreciation method choices. We provide evidence on the construct validity of the Bowen et al. measure (which we term AGG_SCORE) as an indicator of aggressive accounting by showing that the frequency (probability) of negative financial reporting outcomes (restatements with and without irregularities and AAERs) is increasing in AGG_SCORE. We also provide evidence that the Bowen et al. measure is associated with aggressive tax reporting by showing that amounts of unrecognized tax positions recognized under FASB Interpretation No. 48 Accounting for Uncertainty in Income Taxes are positively associated with AGG_SCORE. 2

3 In addition to avoiding the need to identify an external and ex-post indicator of accounting aggression, such as an SEC determination that a GAAP violation has occurred, our use of a simple ex ante aggressiveness score (AGG_SCORE) derived from two common accounting method choices within US GAAP has two related advantages. First, it allows us to consider broad samples over long time periods, and examine the returns to a long-horizon buy-and-hold investment. Second, because AGG_SCORE is derived from legitimate accounting choices, it provides an instrument for aggressive accounting behavior. We do not condition on the existence of an adverse outcome, such as a restatement, so adverse events are far from certain in our sample. Therefore, we can treat an adverse outcome such as a restatement as only one possible outcome of engaging in aggressive accounting behavior. The probabilities of adverse outcomes such as restatements and AAERs are quite small in the sample we consider, although the consequences of these outcomes are severely negative. Specifically, among the most aggressive accounting choice firms, where adverse outcome probabilities are highest, the incidence of restatements is 1.03%, the incidence of restatements coupled with irregularities is 0.30%, and the incidence of AAERs is 0.46%. These results suggest that about 99% of the time, there is no observable negative outcome of aggressive accounting. The combination of a small probability of a large loss suggests the possibility that a rational long-horizon buy-and-hold investor might benefit from holding shares in a firm characterized by aggressive accounting, since the likelihood of the adverse event is low. We probe this possibility by analyzing shareholder returns associated with investments in firms characterized by aggressive accounting over horizons of 3, 5, 10 and 15 years, chosen to span the period of aggressive accounting behaviors, and, possibly, the periods around and after a specified negative outcomes. Our aim is to provide evidence as to whether an investor who holds shares in a firm with aggressive accounting choices experiences lower or higher returns than an investor who holds shares in a firm with less aggressive accounting choices. We also consider the returns to a long-term investment in firms that are characterized not only by most aggressive accounting within GAAP (and therefore captured by our AGG_SCORE measure), but also beyond the boundaries of GAAP. Specifically, we identify firm with both aggressive accounting 3

4 choices and an adverse outcome such as a restatement, and compare those returns to those of investors in similarly ex ante aggressive firms that did not experience an adverse outcome. This analysis is intended to shed light on the cumulative returns to investing in aggressive firms, conditional on those firms experiencing a negative outcome that sharply decreases share values. We test whether the combination of high returns in the periods leading up to an adverse outcome such as a restatement and negative event period returns dominates an investment in a firm without an adverse outcome. We also examine the payoffs to CEOs of firms characterized by aggressive accounting. We define those payoffs as the compensation paid to the CEO over 3-year, 5-year, 10-year and 15-year intervals. We assume that CEO compensation is determined in part by the riskiness of the CEO s strategic choices, including accounting aggression. We expect CEOs and firms to be matched on their aggressiveness at the time a CEO is appointed, that the CEO will rationally demand a risk premium for engaging in more aggressive behavior, including aggressive accounting, and that the board of directors approves of the CEO s choices concerning aggressive accounting. We view our results concerning the association between CEO compensation and accounting aggression as evidence of whether shareholders reward the CEO for the risks associated with aggressive accounting behavior. Our sample for analyzing returns to shareholders consists of all firm-years over 1974 to 2010 with the necessary data to calculate AGG_SCORE. Our tests involving CEO compensation are restricted to firm-year observations on Execucomp between 1992 and Our proxy for accounting aggression, AGG_SCORE, can take on values of 0, 0.25, 0.5, 0.75, or 1. 1 Accordingly, we sort firm-years into five groups, with the least (most) aggressive firms having an AGG_SCORE of 0 (1). In the shareholder returns sample, 58% of firm-years (148,489 of 256,985) are classified in the most aggressive accounting group; less than 4% of firm-year observations (9,061 of 256,985) are classified in the least aggressive group. This disparity in the number of observations has consequences for our research design, since any comparison of the stock returns from these extreme portfolios would have to account for the 16-fold 1 Section 2 describes the calculation of AGG_SCORE and details why it takes on these five discrete values. 4

5 difference in group sizes. To address this issue, we conduct simulations which randomly select investor portfolios with an equal number of firm-years from each of the five AGG_SCORE groups. As noted earlier, our investor returns and CEO payout analyses have horizons ranging between three years and 15 years. For our returns tests, we match monthly stock returns from CRSP to our shareholder returns sample. To be included in this sample, a firm is required to have at least one valid monthly return observation (and an AGG_SCORE). Our database design does not impose a survival bias. Specifically, our data are not truncated by changes in the entity s listing status over time, such as might occur by acquisitions or delistings for other reasons. Rather, we follow each sample firm forward in time and create a continuous series of returns for that firm until the end of our sample period (December 2010). For example, if Firm A is acquired in 1995 by Firm B in a stock transaction, we include the delisting return of Firm A for the acquisition month and use Firm B s return afterwards (in the returns time series attributed to Firm A). We augment Firm A s return series for any subsequent acquisitions. If a firm is acquired for cash or liquidated (or the acquirer is not followed on CRSP or its identity could not be determined), we include the delisted firm s delisting return in the event month and zero returns for every period thereafter. This process ensures that we have a full series of returns data for each firm in our sample with at least one return month. Our simulations randomly choose the investment start month and randomly select firms from each AGG_SCORE group in that month. By design, we require each firm to have a valid AGG_SCORE and to be alive at the time of investment (that is, not yet delisted, not yet acquired). We then measure the buy-and-hold return over the specified horizon of 3, 5, 10 or 15 years. Firms included in each horizon must have returns data over the entire horizon. This condition, coupled with the fact that our returns data end in December 2010, means that the 15-year returns interval includes observations from 1996 and earlier, the 10-year period includes observations from 2001 and earlier, and so on. Our results are summarized as follows. First, we find that the difference in buy-and-hold returns and the difference in average monthly returns, between the most aggressive AGG_SCORE portfolio and the least aggressive AGG_SCORE portfolio are positive over all horizons. The difference in returns is 5

6 monotonically increasing in the length of the horizon. For example, for buy-and-hold returns the differential return to investing in firms characterized by aggressive accounting is about 10% for the 3-year interval, 16% for the 5-year interval, 32% for the 10-year interval, and 52% for the 15-year interval. These differences are all reliably different from zero at the level. When we control for known risk factors (beta, market capitalization and the book-to-market ratio), the returns differential between the least aggressive and the most aggressive firms remains significantly positive for all variations of our model. These results suggest that, all else equal, buy-and-hold investors benefit from investing in firms characterized by aggressive accounting. Second, the difference in total CEO compensation as a percentage of (adjusted) assets between the most aggressive portfolio and the least aggressive portfolio is statistically and economically significant and monotonically increasing over the length of the interval considered. The differential total compensation is about 1% over 3 years, 1.6% over 5 years, 2.7% over 10 years and 3.2% over 15 years. The compensation differential persists after controlling for variables that have been shown to affect CEO compensation (tenure with the firm, age and gender). These results suggest that boards of directors (ultimately, shareholders) reward CEOs who engage in aggressive accounting. Third, we consider how extreme negative outcomes, specifically GAAP violations such as restatements and AAERs, affect returns to shareholders. After we eliminate all sample firms with a future negative outcome, the differences in returns between the most and least aggressive firms actually decreases in magnitude (but remains significantly positive). Using an event time design, we focus specifically on a within-group comparison of the most aggressive firms, where the group of most aggressive firms is divided into those with and those without a negative outcome. We find that negativeoutcome firms have significantly higher returns prior to the negative outcome compared to benchmark firms (those with the same AGG_SCORE and no negative outcome). Confirming results in prior literature, negative-outcome firms experience a significant drop in share price just before and at the announcement of the event. Once the pre-event period is accounted for, however, their overall buy-andhold returns exceed the returns of non-negative-outcome firms. This return differential remains 6

7 significantly positive for more than two years after the public announcement of the negative outcome. Using the existence of a GAAP violation as a proxy for extreme accounting aggressiveness, these findings suggest that long-horizon buy-and-hold investors are rewarded for investing in firms characterized by extreme accounting aggression. Taken together, our results suggest that both long-term investors and CEOs are rewarded for aggressive accounting. In the case of investors, these effects are not eliminated by losses associated with negative outcomes such as restatements and AAERs. That is, long-horizon buy-and-hold investors would be better off, in a returns sense, if the firms they invest in engage in aggressive accounting, even if doing so sometimes results in a negative outcome that engenders a sharp negative market reaction. Once a (long) investor horizon that includes the misstatement period is considered, the overall returns to the shareholders of negative-outcome firms characterized by aggressive accounting are positive on average and higher than returns to other firms. The reason is that the positive cumulative returns prior to the negative outcome outweigh the price decrease at the announcement of the negative outcome. That is, the cumulative returns do not decline to the level of cumulative returns observed for the most aggressive but non-negative outcome firms until about two years following the negative outcome. The rest of the paper is organized as follows. In the next section, we position our research questions in the context of prior work. In Section 3, we describe our sample and data, and provide validity tests of the Bowen et al. construct applied to our sample period, 1974 to Section 4 describes our main empirical tests of the returns to investors and payouts to CEOs. Section 5 extends our tests to consider the influence of negative outcomes that have been linked to aggressive accounting behavior. Section 6 concludes. 2. Background and Motivation In this section, we review results from two literatures which provide a foundation for our analysis. The first of these literatures concerns accounting method choices, and the second investigates market reactions to negative outcomes associated with aggressive accounting. 7

8 2.1. Accounting method choice Research on accounting method choice has largely focused on its determinants, not its consequences. In an early determinants analysis, Skinner [1993] scores three accounting choices (depreciation method, inventory flow assumption, and length of goodwill amortization period) by their income-increasing versus income-decreasing nature, and calculates an overall accounting choices score. 2 He finds that income-increasing choices are associated with firms that are smaller, more levered, and more likely to have an accounting-based bonus plan. These results are not affected by the inclusion of controls for the firm s investment opportunity set. Skinner concludes that a firm s investment opportunity set has only an indirect effect on accounting choice (through debt contracts and bonus plan contracts). Bowen, DuCharme and Shores [1995] also examine determinants of accounting method choice, focusing on inventory valuation methods and depreciation methods. They assign income-increasing methods (FIFO, straight line depreciation) a value of 1, income-decreasing methods (LIFO, accelerated depreciation) a value of 0, and combinations or choices in between (average cost, combination of straight line and accelerated methods) a value of 0.5. The aggregate score, which is an average of the inventory flow assumption score and depreciation method score, ranges between 0 (least aggressive) and 1 (most aggressive), in five discrete groups (0, 0.25, 0.5, 0.75, and 1). Following Skinner, the focus of their study is the determinants of these choices, specifically, the influence of implicit claims, defined as claims arising from arrangements between a firm and its customers, suppliers, employees and short-term creditors. Bowen et al. s results are consistent with their prediction that these implicit claims create incentives for managers to select income-increasing accounting methods. 3 2 Skinner scores income-increasing methods (FIFO, straight line depreciation, and 40-year goodwill amortization period) as a value of 2, and income decreasing methods (LIFO, accelerated depreciation, and a less than 30-year goodwill amortization period) as 0. He treats other methods (average cost for inventory, units of production for depreciation, and goodwill amortization periods of between 30 and 39 years) as neither income-increasing nor income-decreasing, and assigns them a value of 1. 3 Bowen et al. predict and find that one of the implicit claims (employees) may operate in the reverse direction, since firms with large defined pension plans may proxy for the existence of a unionized work force. The existence of a unionized work force may encourage firms to select income-decreasing methods, in the hopes of reducing labor union demands for pay and benefits. 8

9 Finally, Dichev and Li [2006] investigate the relation between sales growth (measured over one year and three years) and aggressive accounting choices, arguing that the success of income-increasing accounting choices (that is, their ability to increase income) depends on growth. For example, the choice between straight-line depreciation and accelerated depreciation will have no effect on income if the balance in gross property, plant and equipment (PPE) does not increase over time. Dichev and Li examine a broader list of accounting choices than previous research; specifically, they examine depreciation method, inventory valuation method, useful life of PPE, full cost versus successful efforts (for oil exploration firms), purchase versus pooling accounting (for firms engaged in acquisitions when pooling was permitted), capital leases versus operating leases, and for defined benefit pension plans, the rate of compensation increase, the expected rate of return, and the discount rate. They find no relation between accounting aggressiveness and their measure of growth and they find no pattern, either consistently income-decreasing or consistently income-decreasing, across accounting choices for their sample. Our study is related to these studies insofar as we consider accounting method choice. Our aim, however, is to analyze some of the outcomes of aggressive accounting, not the determinants of accounting choices. We therefore rely on previous research on the determinants of accounting choices primarily to obtain our measure of accounting aggression. That is, we use Bowen et al. s [1995] measure of incomeincreasing accounting methods as an instrument for aggressive accounting. As pointed out by Dichev and Li [2006], inventory valuation methods and depreciation methods are not intrinsically income-increasing or income-decreasing; rather, their effects depend on increasing balances for these assets and/or increases in input prices for inventories and depreciable assets. Over our long sample period, this assumption seems reasonable given changes in technology and inflation Consequences of aggressive reporting Previous research has documented sharply negative market reactions to announcements of outcomes that indicate a firm has engaged in financial reporting behavior inconsistent with GAAP (that is, accounting practices that are aggressive beyond the boundaries of GAAP). These outcomes include restatements of financial reports, accounting and auditing enforcement releases, and shareholder lawsuits. 9

10 For example, Palmrose, Richardson and Scholz [2004] document an average decline in stock price of 9% over the 2-day window centered on the announcement of a restatement. Their finding that restatements that are associated with fraud have even more negative reactions is consistent with Hennes et al. s [2008] finding that restatements that involve irregularities have a more pronounced negative price reaction than do restatements that involve errors (the 2-day cumulative abnormal return is -14% for irregularities and - 2% for errors). Similarly, Dechow, Sloan and Sweeney [1996] report an 8.8% 1-day decline in stock price at the first announcement of an accounting manipulation that gives rise to an AAER). Hribar and Jenkins [2004] look beyond the immediate price response to restatements to consider the effect of the restatement on the restating firm s cost of equity capital, estimated using observed prices, analysts earnings forecasts and a residual income model. They find a higher cost of equity estimate after a restatement. In the context of our result that firms that restate their earnings experience increases in their share prices in the years leading up to a restatement, and experience price declines at the time of the restatement, it is perhaps not surprising to find that cost of equity estimates increase after restatements. For example, if analysts earnings forecasts do not change, the price decline alone would induce a larger implied cost of equity estimate. Taking Hribar and Jenkins finding at face value, an increase in cost of equity capital is consistent with higher expected returns, since shareholders of such firms would demand a higher return for the incremental risk. Research has also examined the consequences to managers, audit committees, and governing boards of reporting failures. With regard to managers, while Beneish [1999] and Agrawal, Jaffe and Karpoff [1999] find little evidence of increased turnover following fraud and/or GAAP violations, Desai, Hogan and Wilkins [2006] find that about 60% of firms that restate their earnings experience some form of top manager turnover in the two years following the restatement, as compared to 35% turnover for a matched sample of non-restating peer firms. With regard to audit committees, Srinivasan [2005] reports there is greater audit committee turnover (48%) in the three years following a restatement than for a control sample of non-restating firms (35%), and that audit committee turnover is increasing in the severity of the restatement. Fich and Shivdasani [2007] find that following an announcement of a 10

11 financial fraud lawsuit involving a specific firm, directors do not experience abnormal turnover on that board; however, there is a significant decline in other board seats held by these directors. Our work builds on and contributes to this literature in two ways. First, and as discussed earlier, our sample is not conditioned on the existence of a negative financial reporting outcome that has already occurred or will occur with certainty. We focus on an ex ante instrument for aggressive accounting behavior, not an actual occurrence of a negative outcome from such behavior. Our aim is to provide evidence on the association between aggressive accounting practices and buy-and-hold returns for a broad sample over long horizons, not on the association between a negative outcome of aggressive accounting and short window returns associated with that outcome. Focusing on a sample of negative outcomes imposes severe limitations on sample size (negative outcomes are observed for less than 2% of our firmyear observations) and precludes the examination of returns to investing in firms characterized by aggressive accounting but not by negative outcomes. We use negative outcomes in supplemental tests to distinguish among our most aggressive firms and to provide evidence on the returns to investing in these firms over long horizons. We therefore provide a more complete assessment of the consequences of aggressive accounting, by providing a longer-term view of the returns to aggressive accounting than has been previously documented. Second, we extend prior research on the consequences of aggressive accounting for top management. This research has tended to focus on person-specific reputational and career consequences associated with negative financial reporting outcomes such as restatements. In contrast, we associate aggressive accounting with a firm, not any specific person associated with that firm. This presumption is supported by the small number of firm-year observations in our sample with accounting changes that affect AGG_SCORE (i.e., changes in inventory valuation method or depreciation method). In particular, in our sample, 4,083 firm-years or 1.59% of our observations (3,231 firm-years or 1.26%) indicate a change of the inventory valuation method (the depreciation method). Of those, only 1,934 (190) changes do not involve the intermediate choice category, but change directly between FIFO and LIFO (accelerated and straight-line depreciation). Similarly, Dichev and Li (2006) document a very low frequency of 11

12 changes in the depreciation method and inventory valuation method. Thus, the accounting methods we use to characterize aggressive accounting are more of an enduring characteristic of the firm than the outcome of a decision made by any individual affiliated with the firm. 4 We believe this substantial invariance in the accounting methods we analyze suggests that an analysis of payouts to CEOs should abstract from the individual and consider payouts to the CEO position over 3,5,10 and 15-year horizons, regardless of changes in the identity of the person holding the CEO position. 3. Sample and Data 3.1. Sample Construction We construct three datasets for our tests. For our construct validity test, we use a sample of Compustat firms for which we can determine at least one aggressiveness proxy. The returns sample, based on CRSP data, contains firm-month observations with returns. The compensation sample is restricted by the Execucomp coverage. For our base sample, we begin by identifying all firm-years over 1974 to 2010 with total assets of at least $1 Mio. and the necessary data to calculate AGG_SCORE. To calculate AGG_SCORE, we follow Bowen, DuCharme and Shores' [1995] scoring of allegedly incoming-increasing accounting choices. Their approach averages firms choices of two accounting method choices: depreciation method choice (straight line is coded as a value of 1, accelerated depreciation is coded as a value of 0, and a combination of the two is coded as 0.5) and inventory valuation method (FIFO is coded as a value of 1, LIFO is coded as a value of 0; and average cost is coded as a value of 0.5). Given these discrete values, AGG_SCORE can take on only one of five distinct values from 0 to 1, in 0.25 increments. If the necessary data on one 4 Put another way, firms do not seem to adjust their inventory accounting or depreciation methods in response to changes in business fundamentals such as factor input prices, changes in business models, or changes in profitability. This invariance in methods may be associated with earnings attributes. If, for example, in times of rising factor input prices, FIFO and straight-line depreciation tend to increase earnings while the opposite is true in recessionary periods, we might expect more long-term income volatility in FIFO/straight-line depreciation firms. We are exploring this and other possibilities in extensions of our current paper. 12

13 choice is not available, AGG_SCORE consists of the other choice only. 5 In total, 256,985 of the 305,554 firm-years available (or 84%) have the necessary data to calculate AGG_SCORE. Our tests also require us to scale certain variables to control for firm size. We use Bowen et al. s scalar, adjusted total assets. They (and we) measure adjusted total assets as total assets plus LIFO reserve plus accumulated depreciation, depletion and amortization (Compustat data items AT, LIFR, and DPACT). If either the LIFO reserve or cumulative depreciation/amortization is missing, we set the values of these variables to zero rather than delete the observation. Conceptually, adjusted total assets are intended to be a total asset representation that is independent of the depreciation method and inventory flow assumption. We also collect data for the control variables from Compustat: Market capitalization and the book-to-market ratio as of the end of the last fiscal year. If data to compute the market capitalization is missing on Compustat, the market capitalization is taken from CRSP as of the end of the prior month. Beta is estimated over the past five years of monthly returns, ending at the end of the prior month and requiring at least twelve valid returns observations. 6 Data on both depreciation method and inventory valuation method are either do not fall in the respective three categories or are missing for 15.7% of the Compustat sample over the sample period. To probe whether these sample requirements introduce a form of bias, we construct a continuous aggressiveness score (AGG_SCORE2) using the determinants model from Bowen et al. Specifically, we estimate annual cross-sectional regressions of aggressiveness score on a dummy for the firm s membership in a durable goods industry, R&D expense, labor intensity, a dummy for the existence of a defined pension benefit plan, the amount of notes payable, cost of goods sold (interacted with dummy variables for manufacturing and non-manufacturing industries), and advertising expense. Using our sample data with information on AGG_SCORE, we estimate a determinants model using these variables. 5 In the overall sample, 89,075 (3,777) firm-years with data on the depreciation method choice (inventory flow assumption) do not have data for the inventory flow assumption (depreciation method choice). 6 Both requiring beta and the book value of equity will reduce the sample slightly when those control variables are used. Results are not meaningfully affected if these data are required in the sample construction. 13

14 We then use the estimated year-specific coefficients from this model to generate an implied aggressiveness score (AGG_SCORE2) for each firm with data on the determinants, including firms with the original AGG_SCORE. Our sample for tests that use AGG_SCORE2 increases to 279,724 firm-year observations and reduces the sample loss of the Compustat universe to 8.9%. Briefly, we provide evidence on the high correlation of the two measures and obtain very similar results using AGG_SCORE or AGG_SCORE2. Therefore, we report and discuss results for only the former. For our returns dataset, we begin by constructing an augmented CRSP database that uses the delisting and M&A information. Overall, we aim to use a dataset that does not impose any survival bias, and accurately reflects the returns of a passive investor in these stocks. Specifically, our data are not truncated by changes in the entity s listing status over time, such as might occur by delistings for any reason. For each firm in our sample that has at least one valid returns observation on the monthly CRSP database, we track the firm in time to create a continuous series of returns for that firm till the end of our sample period (December 2010). If Firm A is acquired by Firm B in a stock transaction, we include the delisting return of Firm A for the month of the acquisition, and proceed by using Firm B s stock return afterwards. We still attribute the returns time series to an original investment into Firm A. We continue augmenting Firm A s return series using all subsequent acquisitions, using CRSP information to identify the acquirer (data item NWPERM). If a firm is acquired for cash, liquidated, or the acquirer is either not followed on CRSP or could not be determined, we include the delisting return for the event month and a series of zero returns for period thereafter. The latter step essentially reflects a zero interest rate on cash holdings. Overall, this process ensures that for all firms that existed at some point in time within our sample period, we have a full series of returns data till the end of the sample period, December To ensure public availability of the accounting information, we match fiscal year data with stock returns from the returns database above from Month +4 through Month +15 after the fiscal year end. The final dataset ranges from January 1975 to December Our CEO compensation data is taken from ExecuComp, which in turn collects information from annual proxy statements (Form DEF14A). ExecuComp covers mainly concurrent and previous S&P 14

15 1500 constituents. Data prior to 1994 covers mostly the S&P 500 firms. Hence, our sample is reduced considerably, averaging about 1,452 firms per fiscal year with AGG_SCORE available and ranging from 1992 through We examine three compensation variables in separate tests: base salary, total compensation (base salary, cash bonuses, other annual and deferred compensation, long-term incentive plan payouts, restricted stock grants, and stock options grants). We term the difference between the two items other compensation. All compensation variables are expressed as percentages of adjusted total assets (divided by adjusted total assets and multiplied by 100). Industry-adjusted compensation variables are differenced with the corresponding SIC2 industry averages of the same fiscal year. We also use ExecuComp to form our control variables for the compensation tests: Gender is a dummy variable that takes on the value of 1 if the CEO is female. Age is the CEO s age at fiscal year end in years. Tenure is the difference, also in years, between the fiscal year end date and the date of the appointment as CEO Descriptive Statistics Table 2 reports descriptive statistics on key variables in our samples. The average AGG_SCORE is and the median firm has already the maximum value of As apparent from the first column of Table 3, 58% of the sample have the highest AGG_SCORE of 1. The least aggressive group (with AGG_SCORE = 0) has 9,061 firm-years, or only about 3.5% of the total sample observations. The adjusted total assets exceed, by construction, the balance sheet number by about $463 million ($15 million) for the average (median) firm. Descriptive statistics of the returns dataset are on firmmonth observations with a valid AGG_SCORE. Reflective of many firms in our sample that delisted in the past (but continue to have data on Compustat) the median return is 0%. Compensation variables are expressed in percentages of adjusted total assets. For example, the average firm pays % of adjusted assets in total compensation. About 1.8% of our observations have female CEOs. The average CEO tenure (CEO age) is 7 years (55 years). Our tests address the disparity in the number of observations in the AGG_SCORE groups in two ways. First, and easily, we also examine AGG_SCORE2 which is not restricted to five discrete values and does not exhibit a highly uneven distribution: AGG_SCORE2 is cardinal and continuous, so we can 15

16 create equal-size portfolios by assigning all firm-years into AGG_SCORE2 quintiles of 55,945 firm-years each. While this first approach solves the problem of unequal group sizes, it creates (we believe) substantial noise in the data by using estimates of aggressiveness proxy rather than the actual proxy itself. For this reason, we prefer our second approach which retains the focus on AGG_SCORE, but uses a simulation technique to sample from the five AGG_SCORE groups in a manner that ensures that our test samples are of equal sizes. We discuss the simulation design in Section Construct Validity of Aggressiveness Proxies As a first step, it is important to establish the validity of AGG_SCORE as a proxy for aggressive behavior. Towards that goal, we investigate the correlation of AGG_SCORE with several variables that prior research has put forward in one or more contexts and that are intuitively associated with aggressive accounting. Our first aggressive measures relate to negative outcomes experienced by a firm related to their financial reporting; these include a restatement of the financial statements, a restatement where irregularities are noted, and accounting and auditing enforcement releases (AAERs) by the SEC. Data on restatements and irregularities are from Hennes, Leone and Miller [2008]. Data on AAERs are from Dechow, Ge, Larson and Sloan [2010]. For each firm-year in our sample, we count the number of restatements and the subset of these restatements that involve an irregularity within the 11 months following the fiscal year end. We also assess the frequency of accounting and auditing enforcement releases (AAER) by the SEC. Panel A of Table 3 shows these results for each AGG_SCORE group; Panel B shows the same data using the AGG_SCORE2 variable. As is evident from Panel A, the incidence of these negative outcome events is monotonically increasing across the AGG_SCORE groups: the least aggressive groups have the smallest incidence (0.2869% total restatements, % restatements with irregularities, and % AAERs) and the most aggressive group has the highest incidence (1.0331% total restatements, % restatements with irregularities, and % AAERs). Our final validity test examines the relation between aggressiveness in accounting method choice and aggressiveness in tax positions. Our measure of tax aggressiveness is the uncertain tax position 16

17 (UTP). The UTP is the ending balance of a firm's unrecognized tax benefits that is required to be reported under FASB Interpretation No. 48 ( FIN 48 ). FIN 48 requires firms to calculate and disclose their reserves for uncertain tax positions. Uncertain tax positions concern positions (deductions), taken by the firm on their income tax return, where it is viewed as more likely than not that the position taken by the firm will not be sustained upon examination by the income tax authority. That is, uncertain tax positions will need to be reversed by the firm with a probability of 50% or more (and pay incremental taxes). The amount to be recognized in the financial statements is the largest amount that is greater than 50 percent likely of being realized upon ultimate settlement of the tax dispute. We use the ending balance of the UTP, scaled by adjusted total assets, as our measure of the firm s tax aggressiveness. 7 UTP data exist only for the last four years of our sample period and only for a subset of the firms, as reported in Table 1. Table 3 reports the mean and median levels of asset-scaled UTP for each of the five AGG_SCORE groups. As is evident there, the mean and median UTP are generally monotonically increasing across the groups (the exception is the least aggressive group has a slightly higher mean UTP than Group 2). Importantly, the difference in the mean and median UTP between the least aggressive group (mean is 0.52%, median is 0%) and most aggressive group (mean is 1.47%, median is 0.43%) is positive. 4. Main Empirical Tests Our tests in this section focus on the returns to aggressive accounting, as measured by the longterm returns earned by shareholders and the long-term compensation earned by the CEOs of these firms Investor experience simulations 7 Brown, Drake and Martin [2010) as well as Rego and Wilson [2011] also use FIN 48 disclosures of UTP as a measure of tax aggressiveness. Neither study, however, examines the association between UTP and a measure of the firm s financial reporting aggressiveness. Frank, Lynch and Rego [2009] find a significant positive association between tax aggressiveness (proxied by an estimate of the firm s involvement in tax shelters) and financial reporting aggressiveness (proxied by accruals quality). To the best of our knowledge, we are the first study to examine whether UTP is positively correlated with aggressive accounting method choices. 17

18 Our objective in measuring investor experience is to mimic a trading strategy of a passive, diversified, long-term investor in stock who determines his portfolio allocation on aggressiveness. 8 We have no prior as to the length of the investment horizon, other than that the horizon should be long enough for both the aggressiveness to materialize in a good or a bad outcome. Therefore, we explore four different horizons, ranging from three years to 15 years. We aim to provide a realistic return measure for such a long-term investor. We begin by describing our simulation design and then describe our tests of the outcome variables. As noted in Section 3, the simulation use the augmented returns dataset that is intended to be free of any survival bias and actually follows each firm to our sample period end. Because our interest is in comparing the long-term investor experience returns to the most aggressive and least aggressive firms, we require some approach that is resilient to the large disparity in the empirical distribution of the observations in these groups (see Table 3). We select a simulation approach, which allows us to control the number of firms in each group that we examine. Our simulation proceeds as follows. Step 1. We randomize the time of the initial investment as we do not want to impose a timing assumption about the initial investment. To that end, we first select, out of the 432 sample months (January 1975 to December 2010), the subset of months that allows for the entire investment horizon to lie within our sample period. So, for a 15-year horizon, eligible months are January 1996 or earlier; for the 3-year investor experience, all investment months begin in January 2008 or earlier; etc. From the horizon-dependent subset of months, we sample 1,000 months with replacement, whereby the probability of being selected is proportional to the number of sample firms in that month. Step 2. For each start month and each of the five AGG_SCORE groups, we randomly select five portfolios with 15 firms each. For a firm to be eligible for a portfolio, we require it to have AGG_SCORE data from a recent fiscal year, which determines the firm s group and portfolio membership. We also require the firm to still exist in its original listing status. That is, we require the firm to have a valid return 8 To clarify, we are probing a trading strategy and therefore rely on realized returns data only. The tests are specifically not intended and designed to be asset-pricing tests, and they do not use realized returns as proxy for expected returns. 18

19 and rule out circumstances that would make an investment into a specific firm impossible, such as a prior liquidation, takeover etc. 9 Investments are assumed to be made at the beginning of the selected month, with an equal amount ($1) invested in all 15 stocks. Step 3. We compute the buy-and-hold returns for each firm over the investment horizons (3, 5, 10, or 15 years). While we believe buy-and-hold returns are the appropriate measure for our research design, we try to mitigate concerns about the statistical properties of such cumulative returns and any possible resulting bias in our tests by also presenting average equal-weighted monthly returns. Consistent with an equal-weighted initial investment (see Step 1), we average the firm-specific returns for the tests on portfolio-level returns. In addition, control variables for the regressions are also averaged across firms, if available. Throughout the paper, we report t-statistics based on standard errors clustered by time (investment start month) for portfolio-level tests. In the firm-specific tests, we use standard errors clustered by both firm and month. Table 4 reports the grand averages of the portfolio returns, computed across the 5,000 portfolios (5 portfolios for 1,000 months) for each AGG_SCORE group and horizon assumption. Panel A (B) shows the results on cumulative buy-and-hold returns (average returns). We also report the returns difference between the top and bottom AGG_SCORE portfolios, our measure of the differential return to most (versus least) aggressive accounting choice behavior. While not monotonic across all AGG_SCORE portfolios, the results in Table 4 show a consistent, and significant, returns difference between the most aggressive firms (with AGG_SCORE =1) and the least aggressive firms (AGG_SCORE=0). The differential buy-and-hold (cumulative) returns are monotonically increasing over the investment horizon, ranging from about 9.66% for the 3-years horizon, to 16.30% for 5 years, to 32.12% for 10 years, and to 51.89% for 15 years. In all cases, the t-statistics indicate reliable deviations from zero at the level in a two-sided test. Conclusions from the average monthly returns tests, in Panel B, are qualitatively the same. 9 Naturally, we allow for investments in (surviving) acquirer. Requiring the existence of the original firm eliminates the possibility of duplicated investments. 19

20 Our tests so far have not controlled for other factors assumed to explain future returns. In Table 5, we examine the returns differential between the most and least aggressive portfolios (stocks) in portfoliolevel (firm-level) regressions. Specifically, we use known factors as control variables in a regression model that regresses the (future) return variables on beta, firm size (as proxied by the log of market capitalization), and book-to-market (as proxied by the log of the firm s book-to-market ratio). We measure each of these regressors, including AGG_SCORE, as of the investment start month. To be clear, we do not update the firm s AGG_SCORE or control variables after the initial investment has been made, consistent with a passive long-term investor relying on information at the time of the initial investment only. The model we estimate is given by Equation (1): Return t, t+k =α 0 +α 1 AGG_SCORE t +α 2 Beta t +α 3 log(size t )+α 4 log( B M )+ε t t (1) Return is the cumulative buy-and-hold return or the average return over the period t to t+k. t is the month of the initial investment, k is the investment horizon (in months). AGG_SCORE is as defined above. Beta is estimated over the Months t-60 to t-1 using the CAPM model, requiring at least twelve monthly returns observations. Size is the market capitalization as of the most recent fiscal year end from Compustat. If that number cannot be calculated due to missing data, Size is the CRSP market capitalization as of Month t-1. B/M is the firm s book to market ratio, whereby book value is taken from the last fiscal year end, scaled by Size. As we take the log of this ratio as a regressor, B/M will be set to missing for all negative book-value firms. 10 Subscripts for either the firm or the portfolio are suppressed. For the portfolio-level tests, we average beta, size and B/M across the all firms in the portfolio with nonmissing values. If a control variable is missing for all 15 firms, we exclude the portfolio from the analysis that uses the control variable. For example, 101 of the 10,000 portfolios in our simulations contain firms with missing or negative book value data, and are thus excluded in regressions with the book-to-market ratio. Equivalently on the firm level, we drop a firm with missing control variables. 10 The inclusion of these firms in a regression that does not take the natural log of the BM ratio does not change our qualitative results. 20

21 In these test, we focus on explaining the returns differential between portfolios (firms) with an AGG_SCORE of 0 or 1; observations with intermediate values of AGG_SCORE at the time of investment are removed. AGG_SCORE therefore reduces to a dummy variable for the most aggressive firms, such that the coefficient on this variable compares to the returns difference between the most aggressive and the least aggressive firms. For brevity, we tabulate results for only the 5-year investor experience interval; other investor experience intervals yield similar findings. Portfolio-level (firm-level) test results are report in Table 5, Panel A (B). For clarification of our design and as a benchmark, we run the regression on AGG_SCORE only, not imposing further sample restrictions by requiring control variables. Therefore, the coefficient estimates for the AGG_SCORE dummy and the associated t-statistics will be the same as the returns differences in Table 4. Beta, size and book-to-market significantly load in the expected direction, with higher returns for higher-beta, smaller and higher book-to-market firms. Importantly, while the magnitude of the AGG_SCORE dummy decreases when controls are included, it remains economically and statistically significant. This result holds for both buy-and-hold and average returns. On the whole, the results indicate that returns differences due to AGG_SCORE are not explained by, but remain significant in the presence of, beta, size and book-to-market CEO experience tests Our CEO experience tests use a similar simulation approach as described for the investor experience, with three differences. First, the observations here are firm-fiscal years, not months. Second, the sample is restricted to fewer firms with compensation data available and to a smaller sample period, 1992 to Third, because of the decreased number of firms in the base sample, not all AGG_SCORE groups will contain 15 firms per fiscal year. 11 Therefore, we modify the selection approach by sampling the minimum of 15 and the number of firms in the smallest AGG_SCORE group (in that year). In other words, if one AGG_SCORE group has less than 15 firms, we select the minimum number of firms from each group to ensure that the portfolios remain of equal size. At the same time, this procedure entails that 11 There are five fiscal years where this restriction applies (2004 to 2008). 21